Author: Pete

It’s the last week of the month. Your report is due Friday, and all you have is a page of half-sentences, a spreadsheet with this month’s numbers and last month’s report as a vague template. You can get ChatGPT to help you write a report for work in under an hour, but not by typing “write my monthly report”. The way that works: you bring the facts, ChatGPT asks you what’s missing, you agree an outline together, and then it drafts one section at a time while you check every number.

Imagine it’s Thursday at four. Instead of staring at a blank document, you answer eight short questions ChatGPT asked you, approve a one-page outline, and twenty minutes later you’re reading a clean draft that opens with the one thing your manager needs to know. You fix two numbers, change one sentence that doesn’t sound like you, and send it.

Want more practical AI workflows like this one, free? Join Launchpad.

Short answer

Tell ChatGPT who the report is for, what decision or update it has to deliver, and the period it covers. Paste your raw notes and upload your numbers. Then ask it to interview you first: one question at a time, about the facts it doesn’t have. Once the gaps are filled, have it propose an outline, approve or change it, and only then let it write. Draft section by section with the bottom line first, write the summary last, and finish with a check where ChatGPT lists every number in the report next to where it came from.

The order matters. ChatGPT is good at structure, wording and cutting. It does not know what happened in your team this month, and if you let it fill the gaps it will produce something that reads well and is partly made up. OpenAI says this plainly: ChatGPT can sound confident even when it’s wrong, and it should be treated as a first draft, not a final source (OpenAI Help Center: Does ChatGPT tell the truth?).

What ChatGPT is good at in a work report (and what stays your job)

Split the work honestly and the report gets faster and better at the same time.

• ChatGPT does well: turning bullets into clear sentences, suggesting a structure that fits the reader, spotting gaps (“you mention a delay but not its cause”), shortening long paragraphs, making the tone consistent, and doing calculations on an uploaded spreadsheet.
• Stays your job: the facts, the numbers, the judgment calls (“is this a risk or just noise?”), anything political (“do I mention the supplier by name?”) and the final read before it goes out.

If you only have notes from one meeting and need a report out of them, our guide on how to turn meeting notes into a report with ChatGPT is the faster route. This page is for the bigger report you write from several sources: the monthly update, the project report, the quarterly review or the incident write-up.

Gather your raw material first (five minutes)

A report is only as good as what you feed in. Before you open ChatGPT, put these in one place:

• The reader and the goal. “My manager and her boss. They want to know if the project is on track and whether we need more budget.”
• The period. “1 to 30 September.”
• Your notes. Bullets are fine. Messy is fine. Wrong is not.
• The numbers. A spreadsheet export with clear column names works best.
• Last period’s report, if there is one, so the new one matches the format your reader already knows.
• What you want from the reader. A decision, an approval, or just “no action needed”.

Check your company’s rules on putting work data into AI tools before you paste anything sensitive. On a personal ChatGPT account you can turn off “Improve the model for everyone” under Settings, Data controls; OpenAI also says that by default it does not train on content from Business and Enterprise workspaces (OpenAI Help Center: Data controls in ChatGPT). When in doubt, replace client names with “Client A” before pasting.

Step 1: Let ChatGPT interview you

This is the step most people skip, and it’s the one that stops the draft from inventing things. Paste your material and this prompt:

I need to write a [monthly / project / quarterly / incident] report for work.
Reader: [who reads it and what they care about]
Period: [dates]
Goal: [what the report must deliver, e.g. "show we're on track and ask for one extra designer"]

Below are my notes and numbers. Do NOT write the report yet.
First, interview me. Ask me one question at a time about anything the reader will need that is missing, unclear or contradictory in my notes. Stop after 8 questions at most. Never fill a gap yourself; if I don't know an answer, mark it "[to confirm]".

[paste notes here]

Answer each question in a line or two. You’ll notice the questions are often the ones your manager would have asked after reading the report. That’s the point: you answer them before it goes out, not in a reply thread afterwards.

Step 2: Agree the outline before any sentences

Ask for structure first, prose later. Fixing an outline takes thirty seconds; fixing a finished draft with the wrong structure takes half an hour.

Based on my notes and answers, propose an outline for this report.
Give me headings only, with one line under each saying what goes there.
Put the most important message for the reader at the top.
Match the structure of my previous report (pasted below) where it makes sense.

Not sure what shape your report should have? These are common, sensible starting points:

• Weekly or monthly progress report: headline status, what got done, numbers against target, what slipped and why, risks, next period’s plan, what you need from the reader.
• Project report (mid-project or close-out): goal and scope, where it stands against plan and budget, decisions made, open issues, lessons learned, next steps.
• Quarterly review: results against the goals set last quarter, what drove the change, what you’ll do differently, goals for next quarter.
• Incident report: what happened, when, impact, how it was fixed, root cause, what changes so it doesn’t repeat. Keep it factual and blame-free.

Step 3: Draft section by section, bottom line first

One section per message keeps ChatGPT focused and keeps you reading. A long one-shot draft tends to get vaguer the further down it goes.

Write section 1 of the outline: [heading].
Rules:
- Start with the conclusion in one sentence, then the supporting detail.
- Use only facts from my notes and answers. Keep "[to confirm]" markers as they are.
- Short paragraphs. Bullets only for lists of 3+ items.
- Plain words. No "leverage", "synergy", "in today's fast-paced world".
- Max 150 words.

Read each section before you ask for the next. If something’s off, say exactly what: “the delay was caused by the supplier, not by us” works far better than “make it better”.

Step 4: Get the numbers from your spreadsheet, not from memory

For the numbers part of a report, upload the file instead of retyping figures into the chat. ChatGPT can analyze spreadsheets (.xlsx, .csv), summarize trends, build tables and charts, and it runs Python for the calculations. OpenAI recommends clear column headers in the first row, one row per record, and no empty rows splitting the data, and it tells you to review the code and assumptions before relying on the result (OpenAI Help Center: Data analysis with ChatGPT).

Use the uploaded spreadsheet for the "Numbers" section.
1. Compare [this month] with [last month] and with target for: [list the 3 to 5 metrics that matter].
2. Show a small table: metric, this period, last period, target, change.
3. Write 3 sentences on what changed most and the likely reason, based ONLY on my notes. If my notes give no reason, write "[reason to confirm]".
4. Tell me which columns and rows you used for each number.

Point 4 is your safety net. If a total looks odd, you can check the exact rows. One caution from the same help page: exact values from scanned PDFs or images of tables aren’t reliable, so export the real spreadsheet when the numbers matter. If your data only exists as PDFs, our guide to extracting data from PDFs into a spreadsheet gets it into shape first.

Step 5: Write the executive summary last

The summary is the part your reader is most likely to read, so write it when everything else is settled. If they stop after five lines, they should still know where things stand and what you need.

Now write a summary for the top of the report, max 5 lines:
- Line 1: overall status in plain words (on track / at risk / off track) and why.
- Lines 2 to 4: the three facts the reader must know.
- Line 5: what I need from the reader, with a date. If I need nothing, say "No action needed."

Check the report before you send it

A tidy draft is not a correct draft. Run this check every time:

Review the full report. Give me:
1. Every number, date and name in the report, each with the source (my notes, my answer to question X, or spreadsheet rows).
2. Any sentence that states something not in my material.
3. All remaining "[to confirm]" markers.
4. Any place where the tone is vaguer or more upbeat than the facts.

Then do the human part yourself:

• Check three numbers against the spreadsheet with your own eyes.
• Resolve every “[to confirm]”. If you can’t, keep it visible, or ask a colleague before sending.
• Read the summary out loud. If it doesn’t sound like something you’d say in a meeting, rewrite that line.
• Make sure bad news isn’t buried. A slipped deadline is easier to hear from you, in your report, than from someone else later.

Make it sound like you, not like ChatGPT

Readers who know you will notice a report that suddenly sounds like a brochure. Three fixes that help:

• Give it a style sample. Paste a report or long email you wrote yourself and say: “Match the tone and sentence length of this sample.”
• Ban the filler. Add a short list of words you’d never use (“delve”, “robust”, “game-changer”) to your section prompt.
• Keep your own judgment lines. Write the “what I think we should do” sentences yourself. They are the part your manager is really paying for.

Not sure whether ChatGPT or Claude is the better partner for long work writing like this? Our ChatGPT vs Claude for writing guide compares the two on real writing jobs and shows which one fits which task.

When you write the same report every month

If this is a recurring report, don’t start from scratch each time. A ChatGPT project keeps related chats, files and instructions together, and OpenAI lists reusing a project for recurring tasks, including storing datasets and reports and returning each week to generate updates, as one of the ways to use it (OpenAI Help Center: Projects in ChatGPT). Put your report template, a style sample and the five prompts above in the project instructions, then each month upload the new numbers and notes and run the same steps.

If what you send is closer to a short weekly status email than a full report, our guide to automated status updates with Gemini sets that up from a Google Sheet.

Common mistakes

• “Write my monthly report” with no material. You’ll get a confident template full of things that didn’t happen.
• Skipping the interview step. The gaps don’t disappear; ChatGPT fills them with guesses.
• Asking for the whole report in one message. Section by section is slower by two minutes and better by a mile.
• Typing numbers into the chat by hand. Upload the spreadsheet and ask which rows each number came from.
• Writing the summary first. It ends up describing the report you planned, not the one you wrote.
• Sending without your own read. Your name is on it.

Related guides

• One meeting into a report: ChatGPT: turn meeting notes into a report people can act on.
• Someone else’s long report into a short brief: summarize a long report and extract action items with Claude.
• Weekly status emails from a spreadsheet: Gemini for automated status updates.
• Numbers stuck in PDFs: Claude AI extract data from PDF into a spreadsheet.
• Questions about the numbers themselves: use ChatGPT to analyze a spreadsheet for work.
• The follow-up email after you send it: use ChatGPT to draft reply emails faster.
• When the report also has to become slides: how to use AI to make a presentation.
• When the report ends in a hiring request: how to use AI to write a job description.
• When your notes feed into a team member’s yearly review: how to use AI to write a performance review.

Start with this month’s report

Take the report you have due next. Gather the notes and the spreadsheet, paste the interview prompt, and answer the questions. You’ll know after the first one whether this saves you time. For most people, the interview step alone makes the report better, because it catches the questions their reader would have asked.

Want a new practical AI workflow like this every week? Join Launchpad.

The meeting ended an hour ago. Your manager wasn’t there, and she wants “a short write-up” before tomorrow’s leadership call. All you have is a page of half sentences, three arrows and a number with a question mark next to it. The fast way: paste those notes into ChatGPT, tell it who the report is for and what they need to decide, give it the sections you want, and make it mark every gap instead of filling it. You turn meeting notes into a report in about ten minutes, and the checking takes longer than the writing.

Imagine it’s 5:40 pm. You paste your notes, paste one prompt, and a minute later there’s a one-page report on your screen: the bottom line in two sentences, the three decisions, the open risk, and what you need from your manager. Two lines say [CHECK], because your notes never said who owns the budget. You fix those two lines, send it, and close the laptop.

Want more practical AI workflows like this one, free? Join Launchpad.

Short answer

Give ChatGPT four things with your notes: who will read the report, what they need to do with it (approve, decide, stay informed), the sections you want in order, and one hard rule: use only what’s in the notes and write [CHECK] where something is missing or unclear. Ask for the bottom line first. Then read every name, number, date and decision against your notes before you send it.

That’s it. The rest of this page is how to make each of those four parts good, which report shape fits which reader, and how to catch the mistakes ChatGPT makes most often with meeting notes.

A summary and a report are different jobs

A meeting summary is for the people who were in the room. They already know the background. They want the decisions, who owns what, and the deadline, in as few lines as possible. If that’s what you need, our guide on how to use ChatGPT to summarize meeting notes covers it, including action items and the follow-up email.

A report is for someone who wasn’t there. Your manager, a client’s director, a board, the colleague who takes over the project next month. They need three things a summary skips:

• Context. Why the meeting happened and what was on the table.
• Reasons. Not just “we chose supplier B”, but why, and what you gave up.
• The ask. What you need from the reader, by when.

That changes the prompt. If you ask ChatGPT for “a summary”, you get a summary. If you tell it who the reader is and what they have to do next, you get a report.

Before you paste: three quick checks

1. Are you allowed to paste it? Meeting notes often hold client names, salaries, prices or people issues. Check your company’s rules on AI tools first, and take out anything you don’t need for the report. On a personal ChatGPT plan you can turn off “Improve the model for everyone” under Settings, Data controls, so new chats aren’t used for training. OpenAI says it doesn’t train on Business and Enterprise workspace content by default (OpenAI Help Center, data controls).
2. Messy is fine, missing is not. ChatGPT copes well with fragments, abbreviations and arrows. It can’t know what you didn’t write down. If the decision was made in the last five minutes and you stopped typing, add one line about it now, while you still remember.
3. Add three lines of context on top. The meeting’s purpose, who was there (roles, not only names), and who the report is for. Thirty seconds of typing, and it’s the biggest quality jump you’ll get.

How to use ChatGPT to turn meeting notes into a report (step by step)

1. Open a new chat so old conversations don’t leak into this one.
2. Paste your three context lines, then your notes underneath, exactly as they are.
3. Paste the report prompt below and fill in the reader and the sections.
4. Read the [CHECK] lines first. Answer them in the chat (“budget owner is Sam, deadline is 14 November”) and ask for the corrected version.
5. Ask for the gap list: “List every sentence in the report that is not directly supported by my notes.” Delete or fix what comes back.
6. Read it once as the reader. If your manager only reads the first two lines, does she know what happened and what you need? If not, ask ChatGPT to rewrite the bottom line.
7. Paste it where it goes: an email, a Word or Google Doc, or your team’s project page.

The report prompt (copy and adapt)

The two rules that do the most work are the [CHECK] rule and the “decided versus discussed” rule. Without them, a language model tends to smooth your notes into something that sounds complete. A report that sounds complete but turns a suggestion into a decision is worse than no report.

Pick the right report shape for the reader

Swap the six sections in the prompt for the shape that fits. Four common ones at work:

• Internal report to your manager or leadership. Bottom line, decisions with reasons, risks, what you need. Keep it to one screen. This is the default prompt above.
• Client meeting report. Attendees from both sides, what was agreed, open points, next steps per side, next meeting date. Leave out internal opinions and pricing thoughts you wouldn’t say to the client’s face. Ask ChatGPT: “Write this so it can be sent to the client as is.”
• Formal minutes. Date, attendees, apologies, one heading per agenda item, the decision under each, actions with owners. Minutes record what was decided, not every word that was said, so tell ChatGPT to leave out the discussion unless it changed the outcome. If your organization has a fixed minutes template, paste the template into the chat and ask ChatGPT to fill it.
• Handover report. For the colleague taking over: background, decisions so far, who the contacts are, what’s still open, where the files live. Add “assume the reader knows nothing about this project”.

If your report is really a weekly status update built from a tracker or spreadsheet, not from one meeting, our guide to automated status updates with Gemini fits that job better.

Put the bottom line first

Most people write a report in the order the meeting happened. The reader doesn’t care about the order. They want the outcome. The US military’s habit of “bottom line up front” (BLUF) is a simple fix: the first sentence says what was decided and what you need, everything after it is detail (Harvard Business Review, how to write email with military precision).

A bad first line: “On Tuesday the project team met to discuss several topics related to the Q1 launch.”

A good first line: “We agreed to move the Q1 launch two weeks to 3 March, and I need your OK on the extra €4,000 for the agency by Friday.”

If ChatGPT gives you the first kind, reply: “Rewrite the bottom line so the first sentence says the decision and my ask, with the date and amount from the notes.”

Check the report before you send it

ChatGPT is good at structure and wording. It’s not a witness to your meeting. Spend two minutes on these, in this order:

1. Numbers and dates. Compare every one with your notes. A “3” in your notes that becomes “30” in the report is the mistake that costs you most.
2. Names and owners. Is each action owned by the person your notes say? Did it give an owner to something nobody owned? That should be a [CHECK], not a name.
3. Decided versus discussed. Read the decisions list and ask yourself: was each of these really agreed?
4. Tone about people. Make sure the report doesn’t make someone sound like they blocked or failed something, unless that’s both true and needed.
5. The gap list. Run the “not directly supported by my notes” question from step 5 above, even when it looks fine.

When you have a transcript instead of notes

If your meeting tool gives you a full transcript, the same prompt works, but add one line: “First remove small talk and repeated points, then write the report.” A transcript of an hour can be very long, so if ChatGPT drops the end, split it in two halves, ask for the decisions and open points of each half, and then ask for one report from both lists. Our meeting notes summary guide has a section on cleaning messy notes and long transcripts. If you only have the recording, start with our guide on how to transcribe audio with AI, then come back to this prompt.

One report from several meetings

Sometimes the report covers a series: three workshop sessions, or a month of client calls. A ChatGPT project helps here. OpenAI’s Help Center describes projects as a place that keeps related chats, files and instructions together, with project instructions that apply to every chat in it, and suggests them for recurring work like weekly reporting (OpenAI Help Center, projects in ChatGPT).

A simple setup:

• Create a project for the client or initiative.
• Put your report prompt and rules in the project instructions, so you don’t paste them every time.
• Add each meeting’s notes as a file or paste them in a new chat inside the project after every meeting.
• At the end, ask: “Using all meeting notes in this project, write one report with the sections from the instructions. Show which meeting each decision came from.”

The “which meeting” line matters. It lets you check a decision back to the right notes in seconds. The Help Center also lists file limits per plan (for example 5 files per project on Free and 25 on Plus), so on a long series, paste notes into chats instead of uploading every one as a file.

Common mistakes

• Not saying who the reader is. You get a generic summary. One line about the reader fixes most of it.
• Letting it fill the gaps. Without the [CHECK] rule, missing owners and dates get invented. Always keep the rule in.
• Writing in meeting order. The decision and the ask go first. Background goes second.
• Too long. If the reader needs to scroll on a laptop, cut it. Ask: “Shorten to one page without losing any decision or action.”
• Sending it unchecked because it reads well. Smooth writing says nothing about whether the numbers are right. The two-minute check does.

If the raw notes came from Google Meet, you can let Gemini write them during the call and pull action items from the notes doc before you ask ChatGPT for a longer report. Our guide to Gemini for meeting notes covers that earlier step.

Related guides

• Summary and action items for the people who were there: how to use ChatGPT to summarize meeting notes.
• The report you need to read is a long PDF: summarize a long report and extract action items.
• Weekly status from a tracker: Gemini for automated status updates.
• Replies that come back after you send it: how to use ChatGPT to draft reply emails faster.
• A full report built from more than one meeting, with numbers and a reader in mind: how to get ChatGPT to help write a report for work.
• Decisions that were made in Slack, not in the meeting: the best ChatGPT prompt for action items from a Slack thread.
• The meeting agreed on a new way of working that now needs writing down: how to use AI to write SOPs.
• The meeting asked for a deck instead of a document: how to use AI to make a presentation.
• The meeting kicked off a new project that now needs tasks, owners and dates: how to use AI to create a project plan.
• The follow-up meeting needs a tighter plan than the last one: how to use AI to create a meeting agenda.

Start with your last meeting

Take the notes from the last meeting you still owe someone a write-up for. Add the three context lines, paste the prompt, answer the [CHECK] lines, run the gap list. Next time, save the prompt where you can paste it in five seconds, and change only the reader and the sections.

Want a new practical AI workflow like this every week? Join Launchpad.

Want a ready-made template? Try our AI meeting minutes generator prompt for a clean list of decisions and action items.

Recording Teams calls first? Read how to use Otter AI for Microsoft Teams to get a transcript you can paste into this workflow.

You have fourteen supplier invoices sitting in a folder as PDFs, and someone needs them in a spreadsheet by Friday. Retyping them is slow and you will mistype a number somewhere. The faster way: have Claude AI extract the data from the PDFs into one table, with columns you choose, one row per invoice, and a check column that tells you which rows to look at before you trust them.

Imagine it’s Thursday afternoon. You drop the fourteen files into one chat, paste one prompt, and two minutes later there’s a table on your screen with a downloadable Excel file next to it. Two rows are marked “check”, because the totals don’t add up. You open those two PDFs, fix them, and the job is done before your next call.

Want more practical AI workflows like this one, free? Join Launchpad.

Short answer

Upload the PDFs to Claude and tell it exactly which columns you want, in which order, and in which format. Ask for one row per document (or one row per line item), the page number each value came from, and “not found” when a field is missing. Tell it to copy values exactly as printed and not to calculate anything, except one check column you define. Then ask for the result as an Excel or CSV file, and spot-check three rows against the PDFs before you use it.

This works because Claude reads a PDF in two ways at once. Anthropic’s documentation says each page is turned into an image and its text is extracted next to it, so Claude sees the layout of a table and not only the words (Claude PDF support docs). That matters for invoices and statements, where the meaning of a number depends on which column it sits in.

What this is good for (and what it isn’t)

Pulling data out of a PDF is a different job from summarizing it. You don’t want Claude’s opinion. You want the same fields, from every file, in the same shape, so you can sort, filter and add them up. Typical cases at work:

• Invoices, quotes and receipts into a cost sheet (supplier, number, date, net, VAT, total).
• Bank or card statements into a list of transactions.
• Price lists or product sheets from suppliers into one comparison table.
• Order forms, registration forms or timesheets into rows you can import somewhere else.
• Deadlines and owners out of a project plan (there’s a section on that below).

It is not the right tool when the numbers have to be perfect and there is nobody to check them, for example your bookkeeping of record. Use it to do the typing, and keep a person on the checking. If you need to understand a long report rather than lift data out of it, our guide to analyzing long PDF files with Claude fits better. If the numbers you need are in charts, use the guide to analyzing charts in a PDF, because values read off a chart are estimates, not printed data.

Before you upload: four quick checks

1. Are you allowed to upload it? Invoices and statements often hold client names, bank details or salaries. Check your company’s rules on AI tools first.
2. How many files, how many pages? Claude’s Help Center lists up to 20 files per chat and 1000 pages per PDF. Visuals (layout, images, charts) are analyzed in PDFs of 100 pages or fewer; from 101 to 1000 pages Claude reads the text only (Claude Help Center, upload files). For forty invoices, work in two batches of twenty. For a 300-page statement, split it.
3. Is it a scan? A scanned page is a picture. Claude can still read it if it is sharp and straight, but expect more misread digits. Rescan crooked or blurry pages before you start.
4. Is file creation on? If you want Claude to hand you a real .xlsx file, turn on Code execution and file creation in Settings, under Capabilities. Anthropic says it’s available on Free, Pro, Max, Team and Enterprise plans, and on Team and Enterprise an owner can switch it off for the organization (Claude Help Center, create and edit files). Without it you still get a table you can copy.

How to extract data from a PDF with Claude, step by step

Step 1: decide your columns before you open Claude

This is the step people skip, and it’s the one that decides whether the result is usable. Open the spreadsheet where the data has to end up and copy its column headers. Claude should fill your columns, not invent its own. If the sheet doesn’t exist yet, write the headers in the order you’ll want them when you sort.

Step 2: the extraction prompt

Upload the files, then paste this, with your own columns:

I uploaded 14 supplier invoices as PDFs.
Extract the data into one table, one row per invoice, with exactly these columns in this order:

file name | supplier | invoice number | invoice date | due date | currency | net amount | VAT amount | total | page

Rules:
- Copy every value exactly as printed. Do not round, convert or correct anything.
- Write dates as YYYY-MM-DD. Add a column "date as printed" so I can check the format.
- Amounts as plain numbers with a dot for decimals, no currency symbol.
- If a field is not in the document, write "not found". Never guess.
- If a value is hard to read, still fill it in and write "unclear" in a last column called notes.
- Do not calculate any value, except the check column below.

Add one more column called check: net amount + VAT amount minus total. It should be 0.
Show the table, then list every row where check is not 0 or notes is not empty.

The “date as printed” column is there for a reason. 03/04/2026 is the 3rd of April in most of Europe and March 4th in the US. If you don’t make that visible, one supplier’s dates end up quietly wrong.

Step 3: fix the flagged rows

Open only the PDFs Claude flagged and look at the page it named. Most of the time the check column catches one of three things: a discount line Claude didn’t count, a total that includes shipping, or a digit it misread on a scan. Tell Claude what the right value is and ask it to update that row. Don’t ask it to “fix the totals”, or it may change the numbers to make the check pass.

Step 4: get the file

Create an Excel file from the corrected table.
Sheet 1: all rows. Amount columns as numbers, date columns as real dates.
Sheet 2: only the rows that were flagged, with the reason.
Also give me the same table as a CSV file.

Open the file and check that the amounts are right-aligned (numbers) and not left-aligned (text), so your formulas work on them.

Step 5: spot-check before you send it

• Count the rows. Fourteen invoices should give fourteen rows.
• Pick three rows at random and compare every field with the PDF.
• If the documents print a grand total (a statement, a price list), add up your column and compare.

Three minutes of checking is what makes the rest of the work safe to trust.

Tables that run over several pages

Line-item tables, statements and price lists often continue over three or four pages, and that’s where rows go missing: at the page break, or where one line wraps onto two. Ask for this explicitly:

The table of order lines starts on page 6 and ends on page 9.
Give me one row per order line: item code | description | quantity | unit price | line total | page.
A description that wraps onto a second line is still one row.
Skip repeated header rows at the top of each page.
At the end, tell me how many rows you found on each page.

The row count per page is your quick check. If page 7 clearly has 25 lines and Claude says 23, you know where to look.

Extract deadlines from a project plan PDF

The same method works for dates. A project plan, a statement of work or a tender document usually has deadlines spread over several pages, some as real dates and some as “ten working days after approval”. Ask Claude to pull them out into a list you can actually plan with:

From this project plan, list every deadline, milestone and due date.
Columns: task or milestone | owner (if named) | due date (YYYY-MM-DD) | type | depends on | page

type is either "fixed date" or "relative".
For relative deadlines ("two weeks after sign-off"), leave due date empty and copy the wording into depends on. Do not work out the date yourself.
Sort by due date, relative deadlines at the bottom.

Keeping relative deadlines separate is the important part. If Claude turns “two weeks after sign-off” into a date, it has to assume when sign-off happens, and that assumption ends up in your calendar looking like a fact. Once you know the sign-off date, fill those in yourself or give Claude the date and ask it to calculate them. Then paste the table into your planning sheet, or into your task tool if it has a CSV import.

Which AI tool should you use to extract table data from a PDF?

Honest answer: it depends on the PDF.

• Claude is a good fit for mixed documents: different layouts from different suppliers, scans, fields scattered across a page, or when you want rules like “not found” and a check column. It reads both the text and the page image (up to 100 pages per file), and can hand you an Excel file.
• Excel’s own PDF import is worth trying first when you have one clean, digital PDF with a proper table in it. In Excel for Windows you go to Data, Get Data, From File, From PDF, and pick the tables it finds (Microsoft Support, import data with Power Query). No AI, nothing to check for invented values. It isn’t in Excel for Mac, and a scanned page has no real table for it to find.
• ChatGPT or Gemini can do the same kind of job, and the prompts on this page work there too. File limits and features differ per tool and plan, so check their help pages before you upload a big batch.

Whichever you use, the method is what makes it reliable: fixed columns, “not found” instead of guesses, a check column, and a spot-check.

Once the table sits in a spreadsheet, you can start asking it questions. Our guide on using ChatGPT to analyze a spreadsheet for work shows how to get answers you can check against the rows.

Want the totals and lookups to update by themselves next time? Our guide on how to use AI to write Excel formulas shows how to get a working SUMIFS or XLOOKUP for your new table.

Common mistakes

• Letting the AI choose the columns. You get a different table every time and can’t combine batches.
• Asking it to calculate everything. Have it copy what’s printed. Calculations belong in your spreadsheet, where you can see the formula.
• Mixing date formats. Always ask for one format plus the date as printed.
• Uploading forty files at once. The chat limit is 20 files. Work in batches with the same prompt, then merge the CSVs.
• Trusting a clean-looking table. Neat formatting says nothing about whether a 7 was read as a 1. The check column and the three-row spot-check do.

Which Claude PDF guide do you need?

• Data into a spreadsheet: you’re on it.
• Understanding a long report: Claude AI analyze long PDF files.
• Numbers that are only in charts: Claude AI analyze charts in PDF.
• Contracts and legal documents: Claude AI key points from a PDF.
• Study or reading notes: Claude AI turn PDF into notes.
• A report summary with action items: summarize a long report and extract action items.

When it’s the same PDFs every week

If you run this extraction every Monday on the same kind of document, save the prompt somewhere you can paste it from, and keep your column list fixed so every week’s CSV lines up with the last one. When the PDFs arrive by email on a schedule, that’s the point where an automation starts to pay off. Our page on Zapier free plan automation examples is a good place to see what that kind of setup looks like.

If those weekly PDFs are supplier invoices, you can go one step further: our n8n workflow for automatic invoice processing picks them up from Gmail, reads each one into fixed fields, and adds a row to a Google Sheet for you to check.

Would you rather buy a tool than build a workflow? Here is how to choose AI invoice processing software that fits your invoices, and how to test it on a real batch first.

Start with this week’s pile, though. One folder, one prompt, one table. Then decide whether it’s worth automating.

Want a new practical AI workflow like this every week? Join Launchpad.

It’s Sunday evening, the quarterly report PDF has eleven charts in it, and on Monday someone will ask you “so what did revenue do in Q3?” You can scroll and squint, or you can have Claude AI analyze the charts in the PDF for you. The trick is to ask for each chart on its own: its title, its page, the numbers it shows, which of those numbers are printed on the chart and which are read off the axis, and whether the text of the report says the same thing.

Imagine opening your laptop on Monday with a one-page table already done: every chart in the report, its page number, the three numbers that matter, and two lines marked “check this”, because the chart and the text don’t agree.

Want more practical AI workflows like this one, free? Join Launchpad.

Short answer

Upload the PDF to Claude, then work in two passes. First ask for an inventory: every chart, graph and figure, with its page and a one-line description. Then pick the charts you care about and ask Claude to turn each one into a small table of values, marking every value as “printed on the chart” or “estimated from the axis”. Finish by asking it to compare each chart with what the report text claims. Check two numbers yourself before you repeat any of them in a meeting.

This works because of how Claude reads a PDF. Anthropic’s documentation explains that each page is turned into an image and the page text is extracted next to it, so Claude can answer questions about “charts, diagrams, and other non-textual content”, and “analyzing financial reports and understanding charts/tables” is one of the listed uses (Claude PDF support docs).

What Claude can and can’t see in a PDF chart

Three limits decide how much you can trust the answer.

• The 100-page line. Claude’s help center says it analyzes text and visuals (images, charts, graphics) in PDFs of 100 pages or fewer. From 101 to 1000 pages it reads the text only and skips the visuals (Claude Help Center, file uploads). So in a 140-page annual report, Claude may describe a chart from its caption without ever having looked at the bars.
• Small or blurry charts. Anthropic warns that Claude “might hallucinate or make mistakes” with low-quality, rotated or very small images (Claude vision docs). A tiny chart squeezed into a footnote corner is exactly that.
• Values without labels are estimates. If a bar has “4.2M” printed on it, Claude can read it. If it only has a y-axis with gridlines every 5 million, any number Claude gives you is a reading by eye, just like yours would be. That’s fine, as long as you know which is which.

Before you upload: three quick checks

1. Is it a normal, unlocked PDF? Password-protected or encrypted files are not supported by Claude’s PDF reading.
2. Is it over 100 pages? Then cut it. Print only the chart pages to a new PDF (most PDF viewers let you print a page range to PDF), or split the report by section, so each upload stays at 100 pages or under.
3. Can you read the axis labels at normal zoom? If you can’t, don’t expect Claude to. Take a sharp screenshot of that one chart, zoomed in, and upload the image next to the PDF.

How to analyze charts in a PDF with Claude, step by step

Step 1: get the chart inventory

Start with the list. It shows you what is in the file and catches charts you would have scrolled past.

List every chart, graph and figure in this PDF.
For each one give: page number, title (or "no title"), chart type (bar, line, pie, table-like), what it measures, the time period, and one sentence on what it shows.
If you cannot read a chart clearly, say which page and why. Do not guess.

Step 2: turn one chart into a table

Now pick the charts that matter for your question. Ask for one at a time; a single chart per request gets a more careful reading than “do all of them”.

Take the chart on page 14.
Turn it into a table with one row per bar or data point: label, value, unit.
Add a column "source" with either PRINTED (the value is written on the chart) or ESTIMATED (you read it from the axis).
Give the axis range and where the y-axis starts.
Then write two sentences: what changed, and by roughly how much.

The PRINTED or ESTIMATED column is the most useful line in this prompt. It tells you straight away which numbers you can quote and which ones you should say as “around”.

Step 3: check the chart against the text

Reports sometimes say one thing in the paragraph and show another in the graph. Ask for it directly:

For each chart in your inventory, find the sentence in the report text that talks about it.
Quote that sentence with its page number, and say if the chart supports it, contradicts it, or if the text never mentions the chart.
List the mismatches first.

Step 4: spot-check two numbers yourself

Open the PDF, go to the pages Claude cited, and check two values: one PRINTED and one ESTIMATED. If both hold up, you can work with the table. If one is off, ask Claude to read that chart again and tell it what you see. This takes two minutes, and it’s the difference between “Claude said” and “I checked”.

Compare charts across pages or across two reports

A lot of the real value sits between charts: this quarter against last quarter, your region against the total, this year’s report against last year’s. Upload both PDFs in the same chat and ask:

Both files have a chart on monthly active customers.
Put the values from both charts in one table by month.
Flag any month where the two reports show a different value for the same period, and say if the chart scale or definition changed between the reports.

That last line matters. When a metric suddenly jumps, the cause is sometimes a changed definition or a new axis, not a change in the business.

Ask Claude to flag charts that could mislead you

Charts can be accurate and still give the wrong impression. Add this line to your prompt from step 2:

Also tell me if this chart could give a wrong impression: a y-axis that does not start at zero, two different y-axes, cumulative totals shown as if they were per period, missing months, or a different scale than the charts next to it.

A simple example: if a bar chart’s axis starts at 80, a move from 82 to 85 looks like the bar more than doubled. The real change is under 4 percent. Claude can point that out, and then you decide how to say it in your update.

A worked example

Say a supplier sends a 36-page market report with nine charts, and you need two things for a budget call: did prices go up, and is demand in your region growing. The inventory prompt returns the nine charts with pages. Two of them answer your question: a price index line chart on page 8 and a regional demand bar chart on page 21. Step 2 turns each into a small table. The price chart has printed values, so you can quote them. The demand chart has no labels, so its numbers come back marked ESTIMATED, and you say “roughly”. Step 3 shows the summary on page 3 calls demand “strong” while the page 21 chart is flat for your region. That’s the question you bring to the call.

Common mistakes

• Asking “summarize this report”. A general summary tends to skip the figures. Ask for the charts on purpose.
• Uploading the full 150-page file. Above 100 pages Claude reads the text but not the visuals, so the charts are quietly lost.
• Quoting estimated values as exact. If it was read from the axis, say “about”.
• Asking for all eleven charts in one go. You get a quick pass over everything instead of a careful reading of the two charts you need.
• Uploading confidential reports without checking. Make sure your company allows it, and check your data settings in Claude first.

Which Claude PDF guide do you need?

• Charts and graphs (this page): numbers, trends and mismatches inside the figures.
• A very long file with tables and page cites: how to analyze long PDF files with Claude.
• A report you need to brief someone on: summarize a long report and pull out the action items.
• Study or reading notes: turn a PDF into notes.
• A contract or legal document: get the key points from a PDF contract.
• Many PDFs into one spreadsheet: extract data from PDFs into a spreadsheet with Claude.
• Numbers already in a spreadsheet: analyze a spreadsheet for work with ChatGPT.

Turn the chart table into your update

Once your two numbers are checked, ask Claude: “Using the checked table, write a five-line update for my manager: what moved, by how much, the one mismatch, and the question I want answered.” You edit it and you send it. If you write these updates every week, the guide on automated status updates shows how to make it a routine.

Try it on the next report that lands in your inbox. One inventory, two chart tables, two numbers checked by you, and you walk into Monday knowing what the graphs actually say.

Get one practical AI workflow like this at a time, free.

The client’s 22-page agreement landed at 4:50 pm, and they would love it signed by tomorrow. You do not have time to read every clause twice, and you do not want to sign something you skimmed. If you want Claude AI to pull the key points from a PDF like that, ask for the lines you would actually act on: who does what, what it costs, the dates, how you get out, and what could hurt you. Ask for a clause number on every line, then check the ones that matter.

Imagine you open the PDF tomorrow morning and a one-page table is already waiting: your obligations, the payment terms, the renewal date, the notice period, and three clauses marked “ask about this”, each with its page number.

Want more desk-sized AI workflows like this, free? Join Launchpad.

When the PDF is mostly charts and you need the real numbers behind them, not just a summary, see how to use Claude AI to analyze charts in a PDF.

Short answer

Upload the PDF in Claude, then use a key points prompt with a fixed shape: parties, what each side must do, money, dates and deadlines, how it ends, risk, and anything unusual. Tell Claude to cite the clause and page for every point and to say “not found” instead of guessing. Then open the three or four points that would cost you the most if they were wrong and read those clauses yourself.

Claude reads both the text and the visuals of a PDF, so tables and signature pages count too. Anthropic’s own PDF documentation lists “extracting key information from legal documents” as one of the main uses (Claude PDF support docs).

Key points are not a summary

A summary tells you what a document is about. Key points tell you what you have to do, pay, watch, or question. That is a different job, and it needs a different prompt.

• Key points you act on (this page): obligations, money, dates, exit terms and red flags, each tied to a clause. Best for contracts, policies, terms of service, NDAs and leases.
• A report summary with action items: for a board pack or a research report you need to brief someone on. See how to summarize a long report and extract action items with Claude.
• A short summary in ChatGPT: when you work in ChatGPT and need a brief for your manager or team from a long policy, tender or proposal. See how to get ChatGPT to summarize long documents for work.
• Notes you keep and study: outline, glossary and cue questions for a manual or training PDF. See how to turn a PDF into notes with Claude.
• Deep analysis of a long file: tables, charts and very large documents. See how to analyze long PDF files with Claude.
• Many PDFs into one spreadsheet: invoices, statements or price lists into fixed columns, one row per document. See how to extract data from a PDF into a spreadsheet with Claude.

Before you upload: three quick checks

1. Can you share it? A contract often carries a confidentiality clause of its own. If you are on a free or personal Claude plan, check your privacy settings first. Anthropic says chats on Free, Pro and Max can be used to train its models when that setting is on, and that Team, Enterprise and API use fall under its commercial terms instead (Anthropic consumer terms update). When in doubt, ask whoever owns the document.
2. Is the file readable? Password-protected PDFs will not work. A blurry scan gives blurry answers. If you can select the text in your PDF viewer, you are fine.
3. Is it the final version? Key points from last week’s draft are worse than none. Check the date or version number on the first page.

How to get the key points from a PDF with Claude

1. Open a new chat at claude.ai. Click the “+” button in the lower left of the chat box, choose “Add files or photos” and pick your PDF. Dragging the file into the chat works too (Claude Help Center: upload files).
2. Paste the prompt below. Fill in who you are in the deal, because “key” depends on which side you sit on.
3. Read the table, not the friendly paragraph around it. Mark every row you are unsure about.
4. Ask the follow-up questions further down this page.
5. Check the clauses that matter most, then write your reply.

The key points prompt

Copy this, change the bits in brackets, and keep the structure. The structure is what stops Claude from drifting into a generic summary.

I am [the freelancer / the buyer / the tenant] in this document.
Pull out the key points I need to act on. Use a table with
these columns: Topic | What it says in plain English | Clause and page.

Cover, in this order:
1. Parties and what the document is for
2. What I must do (deliverables, duties, standards)
3. What the other side must do
4. Money: amounts, when it is due, late fees, expenses
5. Dates: start, end, renewal, every deadline and notice period
6. How it ends: termination rights for each side, and what happens after
7. Risk: liability caps, indemnities, penalties, insurance
8. Ownership and confidentiality: who owns the work, what I cannot share
9. Anything unusual or one-sided compared with a typical document of this kind

Rules: cite the clause number and PDF page for every row.
If something is not in the document, write "not found". Do not guess.
End with the 3 points I should question before I sign.

Why the “who I am” line matters: a 60-day payment term is a key point for a freelancer and a footnote for the buyer. Telling Claude your side changes what it flags.

Extract key points from a legal PDF with Claude

Contracts, NDAs, leases and terms of service hide the important parts in predictable places. When the PDF is a legal document, add these lines to the prompt:

• “Quote the exact wording for rows 4 to 7.” Money, dates, termination and risk are where a paraphrase can quietly change the meaning. A quote lets you check it in seconds.
• “List every defined term that changes what a clause means.” “Deliverables”, “Confidential Information” and “Business Day” are often defined on page 2 and used on page 14. The definition is the clause.
• “Does anything renew automatically? How and when do I stop it?” Auto-renewal with a notice window is one of the easiest things to miss.
• “Which law and which courts apply?” Usually one line near the end, and it decides where a dispute would be handled.
• “Write the questions I should ask the other side, and the ones for a lawyer.” This turns the key points into a short to-do list instead of a worry.

Keep the expectation honest. Claude can tell you what the document says and point at clauses that look unusual. It cannot tell you whether a clause is enforceable where you live, and it is not legal advice. For anything with real money or real liability, use the key points to make your lawyer’s half hour count, not to skip it.

Still choosing between Claude and ChatGPT for contracts? Our ChatGPT vs Claude on Reddit guide compares how much each one can read at once and what legal and accounting users report.

Follow-up questions that catch what the first pass missed

The first answer is a draft. These questions find the gaps:

• “Put every date, deadline and notice period in one table, sorted by date.”
• “What happens if I deliver late, or if they pay late? Quote the clause.”
• “Which obligations continue after the contract ends?”
• “Is there anything in the appendices or schedules that changes the main text?”
• “Rewrite the three riskiest clauses in plain English, one sentence each.”

If Claude answers with a page number that does not match, say so. Ask it to use the page numbers your PDF viewer shows, not the numbers printed in the footer. Anthropic’s help center gives the same tip.

Check the key points against the clauses

This is the step that makes the table trustworthy. It takes about as long as making a coffee.

1. Pick the rows that would hurt most if they were wrong: payment, termination, liability and anything marked unusual.
2. Open each clause and read it yourself, including the definitions it uses.
3. If a row says “not found”, search the PDF for one or two words yourself. Sometimes the clause is there under a different heading.
4. Fix the table, not just the chat, so the version you keep is the checked one.

Long PDFs, scans, and two versions of the same contract

• Length: according to Claude’s help center, Claude reads text and visuals in PDFs of up to 100 pages, and text only from 101 to 1000 pages. If your signature page or price table is an image inside a longer file, upload those pages as a separate PDF.
• Several files: you can attach up to 20 files to one chat. Upload the main agreement and its schedules together and ask Claude to treat them as one document.
• Two versions: upload the old and the new PDF and ask: “List every change between version 1 and version 2 that affects money, dates, obligations or risk. Cite both clause numbers.” It is no substitute for a proper redline, but it shows you where to look first.

When to stop and get a person

Use Claude for speed, and a human for judgment, when:

• the liability or penalty amounts are bigger than you could comfortably pay,
• the document gives away ownership of your work, your data or your clients,
• the key points table has more than two rows marked unusual or one-sided,
• or you would simply sleep better if someone qualified had read it.

That is not a failure of the tool. It is the tool doing its job: getting you to the right questions faster.

If contracts land on your desk every week, a dedicated tool can be worth paying for. Here is how to choose AI contract review software and test it on your own agreements first.

Turn the key points into your reply

Once the table is checked, ask Claude for the email: “Using the 3 points to question, write a short, friendly reply to the client asking to change clause 9.2 to 30-day payment and to cap liability at the contract value.” You edit it, you send it. If you send a lot of these, the client communication templates and the guide to proposal writing for freelancers cover the messages before and after the contract.

Try it on the next PDF that lands in your inbox. One prompt, one checked table, and you go into the call knowing exactly which three clauses you want to talk about.

Get one practical AI workflow like this at a time in the free Launchpad.

Not sure which tool to read documents with? See our guide to AI document analysis tools for work.

You downloaded the PDF. You meant to read it. Then the week got loud, and now that file is still sitting in Downloads while you vaguely remember “something about compliance” or “a process change.” If you want Claude AI to turn a PDF into notes you can actually keep, start by asking for a reusable notes set, not a one-line summary.

Imagine it is Monday morning, you open Notion, and the 40-page training PDF from Friday is already an outline, a short glossary, and ten cue questions you can quiz yourself with before the team call.

Want free, practical AI desk tips like this? Join the free Launchpad on robotshelp.me.

When the PDF is mostly charts and you need the real numbers behind them, not just a summary, see how to use Claude AI to analyze charts in a PDF.

Short answer

Upload the PDF in Claude (claude.ai), then ask for notes in a fixed shape: section outline with page cues, key terms, and open questions you can study later. Copy the result into Notion, OneNote, or Word. Spot-check two or three claims against the source pages before you trust the file at work.

Claude can read PDF text and images (including charts and tables) when you upload a standard, unlocked file. Anthropic’s docs list request limits around 32 MB and up to 600 pages per request (fewer when the context window is smaller), and they recommend splitting dense or oversized files. Details: Claude PDF support documentation.

Notes, summary, and analysis are different jobs

People say “summarize this PDF” when they really want three different outcomes:

• Notes you keep (this page): structured outline, glossary, cue questions, a format you revisit for weeks.
• One-off summary + action items: useful for a report you need to act on once. See how to summarize a long report and extract action items with Claude.
• Deep analysis: tables, charts, page-cited answers, and splitting huge files. See Claude AI analyze long PDF files.
• Key points from a contract: obligations, payment terms, dates and exit clauses you need to act on before you sign. See Claude AI key points from a PDF.

If your input is raw meeting notes rather than a PDF, that is a different workflow. For ChatGPT on meeting notes, start with how to use ChatGPT to summarize meeting notes.

Prep the PDF before you upload

Two minutes of prep saves a messy notes dump.

1. Use a normal PDF (not password-protected or encrypted).
2. Prefer a text-based file over a scan. If every page is a photo of a page, expect more gaps.
3. If the file is huge or dense, split by chapter or section first. Anthropic notes that dense PDFs can fill the context window before you hit the page limit.
4. Name the file clearly, e.g. onboarding-module-3.pdf, so your prompt can refer to it.

How to use Claude AI to turn a PDF into notes

1. Open Claude and start a new chat.
2. Upload the PDF (paperclip / attach).
3. Paste a notes prompt (templates below).
4. Ask Claude to keep page numbers next to each major point so you can verify later.
5. Copy the output into your notes app. Do not leave the only copy inside the chat.

If Claude truncates or skims, say: “Continue from section 4. Keep the same notes format.” Or upload the next chapter as a separate file and merge the notes yourself.

Prompt: structured outline notes you can reuse

Copy this after you attach the PDF:

Turn this PDF into reusable study notes I will keep in Notion.

Format exactly:
1) Document snapshot (3-5 lines: what it is, who it is for, why it matters at work)
2) Outline notes
   - H2-style section titles from the document
   - Under each: 3-7 bullet points in plain language
   - Add (p. X) when you can see a page number
3) Key terms glossary (8-15 terms). Each line: Term: short definition in my words. If you cannot find a page, say "page unclear".
4) Cue questions (8-12 open questions I can use as flashcards, with short answer keys)
5) Gaps: anything important that was unclear, scanned, or cut off

Rules:
- Do not invent facts, numbers, or policy rules.
- Prefer concrete verbs and examples over vague summaries.
- If two sections repeat, merge them and say so once.

That prompt is the core of “claude ai turn pdf into notes”: you get a notes pack, not a chatty essay.

Want to see the whole document on one page? Paste those outline notes into the steps from our guide on how to make a mind map with AI and you get a branching map that shows how the sections connect.

Prompt: glossary and cue questions only

When you already have an outline and only need study aids:

From this PDF, build only:
A) Glossary: 12 key terms with one-sentence definitions in plain English (include page if visible)
B) Cue / flashcard questions: 10 questions that test understanding, not trivia
   For each: Question / Short answer / Page cue

Skip long summaries. Flag anything you could not verify in the file.

Claude create study notes from a training manual PDF

Training manuals are where this shines. You do not need a literary summary. You need something you can revise from before a quiz, a go-live, or a client call.

Try this after uploading the manual:

This is a workplace training manual PDF. Create study notes for someone who must apply it on the job next week.

Include:
1) Learning outcomes (what I should be able to DO after studying)
2) Step-by-step procedures (numbered; keep tool names and thresholds exactly as written)
3) Common mistakes / warnings called out in the manual
4) Mini scenarios: 5 "what would you do" questions with answer keys grounded in the text
5) Quick reference: a short checklist I can print

Do not soften safety or compliance language. Quote critical rules tightly and mark the page.

Work through one module at a time if the manual is long. Your notes stay sharper, and you avoid a half-finished mega-dump.

Export notes to Notion, OneNote, or Word

Claude gives you text. Your system of record should be your notes app.

• Notion: paste as Markdown if Claude used headings and bullets; turn the glossary into a simple table later if you like.
• OneNote: paste into a section named after the course or policy; keep one page per chapter.
• Word: paste, then apply Heading 1 / Heading 2 so you can generate a table of contents.

Add a line at the top of every notes file: source filename, date you generated the notes, and Claude chat date. Future-you will thank you when a policy updates.

Check your notes against the source pages

AI notes are a draft until you verify them. Spend five minutes:

1. Pick three claims that would embarrass you if wrong (a number, a deadline, a required step).
2. Jump to the cited pages in the PDF.
3. Fix wording in your notes file, not only in the chat.
4. If Claude said “page unclear,” find it yourself or delete the line.

This is the difference between “I asked Claude” and “I have notes I trust.”

When the PDF is too long for one pass

If the upload fails, times out, or the notes feel thin:

• Split by chapter and run the same prompt on each part.
• Ask for notes on pages 1-40, then 41-80, and merge.
• For scanned books, improve clarity first (cleaner scan, OCR) before another upload.

For chart-heavy or research-style reading where you need page-cited extraction more than study notes, switch to the longer-PDF analysis guide linked above.

If those notes are for a course or an exam, the next step is to make Claude quiz you on them. Here is how to use Claude to study.

A simple weekly habit

Pick one PDF you already meant to learn (onboarding doc, vendor guide, internal playbook). Run the outline-notes prompt once. Put the result in your notes app the same day. Next week, use only the cue questions for a ten-minute refresh. That loop is how PDF notes become work memory, not another abandoned chat.

If you want more desk-sized AI workflows like this, the free Launchpad on robotshelp.me is the starting point.

You drop a 180-page PDF on your desk and need answers by lunch: what the tables say, what the chart on page 47 shows, and which pages back each claim. A plain “summarize this” prompt will not get you there. This guide shows how to use Claude AI analyze long PDF files the practical way: structured questions, page cites, table and chart reads, and a clean split when the file is too big for one pass.

Imagine it is 11:15. Your PDF is still open. You already have a one-page brief with numbers from Table 3, a note on the Q3 chart, and three quotes with page numbers you can paste into Slack without guessing.

Want one practical AI workflow like this every week? Start free at RobotsHelp Launchpad.

When the PDF is mostly charts and you need the real numbers behind them, not just a summary, see how to use Claude AI to analyze charts in a PDF.

Short answer

Upload a standard PDF (no password) to Claude, ask for structured output with page numbers, and treat tables and charts as first-class sources. Anthropic’s PDF support docs state that Claude reads both the extracted text and a page image, so charts and tables are in scope. Limits on a request are up to about 32MB, and up to 600 pages (or 100 when the context window is under 1M tokens). Dense files can fill context earlier, so split them. If you mainly need a short brief and a task list from a report, use our guide to summarize a long report and extract action items instead. This page is for analysis: compare, extract, cite, and check.

When this helps (and when it does not)

Use this workflow when you need to:

• Pull numbers from tables and check them against the narrative
• Read a chart or figure and say what it shows in plain language
• Answer a specific question with page cites you can defend in a meeting
• Extract fields into a fixed structure (risks, owners, dates, amounts)

Skip it when the file is a scanned mess with no readable text and no clear layout, when the PDF is password-locked, or when you only need a short summary of meeting notes. For notes, ChatGPT for meeting notes is the better starting point. Spreadsheet and CSV analysis is a different job; keep that for its own page.

Prep the file before you upload

Five minutes of prep saves a bad answer.

1. Unlock and flatten. Claude needs a standard PDF with no password or encryption (same requirement in Anthropic’s docs).
2. Check size and pages. If the file is huge or page-dense, plan a split by chapter, section, or page range before you start.
3. Name your question. Write one sentence: “I need X from this PDF so I can do Y.” Vague uploads get vague answers.
4. Note the pages that matter. If you already know Table 4 and Appendix B matter, say so up front.
5. Keep a clean original. Do not ask Claude to “fix” the PDF. Ask it to analyze, then you edit the real file offline if needed.

If you are choosing which model to run for a heavy document job inside Grok Bot, our note on picking the best model per task helps you match the tool to the job.

How to use Claude for long PDFs: step by step

Here is the desk path that matches how people actually search how to use Claude for long PDFs.

1. Open Claude (claude.ai or your work workspace) and start a new chat.
2. Attach the PDF. Prefer one clear file over a zip of scans.
3. Put your question and output format in the same message as the file. Anthropic recommends placing the document before the text instructions when you control the request order.
4. Ask for page numbers on every claim. Use the page numbers you see in the PDF viewer.
5. Ask a second pass for tables and charts only: “List every table and figure, what each shows, and the page.”
6. Spot-check two numbers and one chart yourself before you forward anything.

Claude processes each page as text plus an image. That is why it can talk about a bar chart that never appears as plain text. It is also why a dense pack of small-font pages can burn context faster than a light 40-page memo.

First time using Claude at all? Start with our Claude tutorial for beginners for the sign-up, your first work prompt and the + button for adding files, then come back here for the long PDF.

Ask for tables and charts on purpose

Do not assume a general summary will catch the figure on page 62. Call it out:

You are helping me analyze this PDF for work.

1) List every table: title, page, and the 3 most important numbers.
2) List every chart or figure: title, page, and what it shows in one sentence.
3) Flag any place where the text and a table disagree.
4) Cite page numbers for every point.

If something is unreadable, say "unclear on page N" instead of guessing.

Demand page-cited answers

For a decision brief, force the cite format so you can verify fast:

Answer my question using only this PDF.

Question: [paste your question]

Format:
- Direct answer (max 5 lines)
- Evidence bullets, each ending with (p. X)
- Open questions the PDF does not answer

Do not invent figures. If the PDF is silent, say so.

Structured extraction prompt you can reuse

When you need fields, not prose, lock the shape. Recruiters skimming a resume PDF can use a short version of this; the same pattern works for contracts, vendor packs, and board packs.

Extract from this PDF into this exact structure.

## Document
- Title:
- Date:
- Author / org:

## Key facts
- Fact | page

## Numbers that matter
- Metric | value | page | notes

## Risks / issues
- Issue | severity (low/med/high) | page

## Decisions needed
- Decision | owner if named | page

## Quotes worth keeping
- "quote" (p. X)

Rules: no invented numbers, mark missing fields as "not stated", keep each line short.

Save the filled block in your notes. Next week, paste a new PDF and the same structure. That is how analysis stays comparable across documents.

If you want notes you keep and come back to, rather than one analysis, turn the PDF into an outline and glossary with Claude instead.

If the PDF is a contract, an NDA or another agreement, use our guide to extracting key points from a legal PDF with Claude to get a short list of who does what, what it costs, the dates and the risky clauses, each with its page.

If the job is data entry across many PDFs (invoices, statements, price lists) into one spreadsheet with fixed columns, use our guide to extracting data from a PDF into a spreadsheet with Claude.

When the PDF is too long for one upload

If Claude truncates, times out, or starts skipping middle sections, split the work. Anthropic notes that dense PDFs can fill the context window before you hit the page cap, and that splitting into sections helps.

1. Split by logical parts: executive summary, body chapters, appendices.
2. Run the same structured prompt on each part.
3. In a fresh chat, paste the structured outputs only (not the whole PDF again) and ask for a merge: conflicts, duplicates, and a final brief with page cites preserved.
4. If images inside the PDF are huge, exporting a lighter PDF (or downsampling) can help, because each page is also handled as an image.

Do not merge by asking for “a nicer summary of everything you said.” Merge by asking for conflicts and missing cites. That keeps you honest.

If the long PDF is in another language, translate it first and analyze the translation. Here is how to use AI to translate a document without losing your terms or the layout.

How to analyze a long PDF with ChatGPT

Same job, different tool. OpenAI’s Chat with PDFs use case is built for asking questions across uploaded documents. The desk pattern is close to Claude’s:

1. Upload the PDF in ChatGPT.
2. State the question and the output shape in the first message.
3. Ask for page references and a separate pass on tables and figures.
4. If the file is long, split by section and merge the structured notes.

Use ChatGPT when your team already lives there, or when you want a second opinion on the same extraction. Use Claude when charts and layout-heavy pages are the main risk. Many people run both on the same file for a high-stakes number: if the two answers disagree, go back to the page yourself.

I uploaded a long PDF. Analyze it for work.

Give me:
1) A 8-line brief
2) A table of key metrics with page numbers
3) A list of charts/figures and what each shows
4) Three risks or gaps, each with a page cite

If you cannot read a page, say which page and why.

If you only need a short, checkable summary of a long work document rather than a full analysis, follow how to get ChatGPT to summarize long documents for work: map the file first, summarize it section by section, then write the brief for the person who will read it.

Quality check before you send anything

Treat the model like a sharp intern, not a source of truth.

• Open two cited pages and confirm the quote or number sits there.
• Re-read one chart yourself. Ask: would I bet my name on this sentence?
• Search the PDF for a unique phrase from the answer. If it is not there, fix or drop it.
• Keep “not stated” results. Empty fields are better than invented ones.

If the answer feels smooth but cites are thin, re-run with a harder rule: every bullet must end with a page number or the word “not stated”.

Common mistakes

• Only asking for a summary. Summaries hide the table you needed. Ask for structure.
• No page cites. You cannot defend the brief in a meeting without them.
• One giant upload with a vague prompt. Name the question, the audience, and the format.
• Ignoring charts. If the decision sits in a figure, say so in the prompt.
• Forcing a password PDF through. Unlock it first, or the run fails the requirements.
• Skipping the spot check. One wrong number in Slack travels farther than a careful correction.

Quick checklist

• Standard PDF, unlocked, question written in one sentence
• Prompt asks for structure + page cites + tables/charts
• Long file split by section if needed, then merge structured notes
• Two cites and one chart verified by hand
• Brief ready to paste, with “not stated” left visible where the PDF is silent

Write your first analysis today

Pick one real PDF from this week. Run the structured extraction prompt once. Check two pages. That single pass teaches you more than another generic “summarize this” try. Want more workflows like this, one practical win at a time? Join free at RobotsHelp.

Comparing options for long files? Our guide to AI document analysis tools shows how to pick one.

You open ChatGPT, type “write a cold email to a potential client”, and get back something you would delete in two seconds if it landed in your own inbox. The prompt is the problem, not the tool. The ChatGPT prompts for cold email outreach below work in a fixed order: write your offer down once, find one true reason to contact this person, draft a short email with hard rules, then let ChatGPT read it as the prospect before you hit send.

Imagine it is 8:40, your first call is at 9, and five cold emails are already written. Each one opens with something true about that person’s business, each one asks one small question, and not one of them sounds like a template.

Want one practical AI workflow like this every week? Start free at RobotsHelp Launchpad.

Why most AI cold emails get ignored

A generic prompt gives a generic email. ChatGPT does not know who you help, what you have actually done, or why this person and not the next one. So it fills the gaps with the safest words it has: “I hope this finds you well”, “I came across your company”, “we help businesses like yours grow”. Your prospect has seen those lines a hundred times this month.

OpenAI’s own prompt engineering best practices say the same thing in plainer terms: be specific about context, outcome, length, format and style, replace vague wording with exact limits, and say what you want instead of only what you do not want. Every prompt on this page follows those rules.

This page is about the first email: the one that has to earn a reply from a stranger. If you need the follow-ups after it, our guide to ChatGPT for sales outreach sequences builds the full three-step plan.

Step 1: Write your offer card once

Before any prospect, give ChatGPT the facts about you. Save the answer in a note. You will paste it at the top of every cold email chat from now on, so you never start from a blank page again.

Help me write a short "offer card" I can reuse for cold emails. Ask me these questions one at a time, then write the card:
1) Who exactly do I help? (role, type of business, size)
2) What problem do they have that I fix?
3) What do they get, in one sentence, if they work with me?
4) One proof point that is true: a past client result, a portfolio piece or a number I can back up.
5) The smallest next step I could offer (a 15 minute call, a short audit, a sample).

Format: 5 lines, plain words, no adjectives like "innovative" or "cutting-edge". If my answer is vague, ask me for a concrete example before you write.

The fifth line matters most. A stranger rarely says yes to a sales call, but many will say yes to something small and useful, like “want me to send the three fixes I would make to your booking page?”

Step 2: Find one true reason to write to this person

Personalization is not “I love your website”. It is one specific thing about them that connects to the problem you fix. Copy the text from their About page, a recent LinkedIn post, a job ad or a product page, and paste it in. Do not let ChatGPT browse and guess: you choose the source, so you know every fact is real.

Here is my offer card: [paste card]

Here is public text about my prospect, [name, role, company]:
"""
[paste About page, recent post, job ad or product page]
"""

From this text only, find up to 3 possible reasons to contact them that connect to my offer.
For each reason: quote the exact sentence it comes from, and say in one line how it links to the problem I fix.
If nothing in the text connects to my offer, say "no clear reason" instead of making one up.

“No clear reason” is a useful answer. It saves you from sending an email that would have been ignored anyway, and it tells you to skip that prospect or look for a better source.

Step 3: Draft the first email with hard rules

Now the actual email. The rules in this prompt are what turn ChatGPT from a brochure writer into something that sounds like a person typing between meetings.

Write a cold email from me to [name, role] at [company].
Use my offer card and reason number [1, 2 or 3] from above.

Rules:
- Under 90 words in the body.
- First line: the specific reason, in plain words. Not a compliment.
- Second part: the problem it points to and what I do about it, in two sentences max.
- Use my proof point only as written. Do not add numbers, client names or results.
- End with ONE small question they can answer in a few words.
- No "I hope this finds you well", "just reaching out", "I came across", "quick question", "synergy".
- No exclamation marks. Write like a calm professional, not a marketer.

Output: the email body only, in plain text.

Read the draft out loud. If one sentence sounds like something you would never say on the phone, ask for a rewrite of that sentence only: “Rewrite line 2 the way I would say it to a colleague.”

How to make ChatGPT write a cold outreach email to a potential client

If you are a freelancer or consultant writing to a potential client, not a sales rep with a CRM full of leads, the email is less “here is our product” and more “here is something I noticed, and here is how I can help”. This variant puts your work up front.

I am a [freelance designer / bookkeeper / consultant]. Write a cold email to [name], who runs [business].

What I noticed: [one specific, true observation, e.g. "their booking page has no prices and the form asks for 9 fields"].
What I could do: [the small next step from my offer card].
Relevant work: [one past project in one line, only if it is real].

Under 80 words. Start with what I noticed. Offer the small next step for free or as a first conversation, in my words. Finish with a yes or no question.
Then give me a second version that is even shorter, under 50 words.

Two versions side by side make it easy to pick. Many freelancers find the shorter one wins more often than they expect, because a busy owner can read it on a phone between two jobs.

Step 4: Subject lines that are short and honest

The subject line has one job: get the email opened without tricking anyone. That is also a legal point. The FTC’s CAN-SPAM compliance guide says the subject line must accurately reflect the content of the message, and Google’s email sender guidelines warn against starting a subject with “Re:” or “Fwd:” unless the message really is a reply or a forward.

Write 8 subject lines for the email above.
- 2 to 5 words each, lowercase is fine.
- Each one must honestly describe what the email is about.
- No "Re:", no "Fwd:", no fake urgency, no clickbait questions.
- Mix: 3 that name their company or project, 3 that name the problem, 2 that name the small offer.

Step 5: Let ChatGPT read it as the prospect

This is the step most people skip, and it catches the weak spots before a real person does. Open a fresh chat so ChatGPT has not seen your reasoning, and give it only what the prospect would see.

You are [role] at [type of company]. You get about 30 sales emails a week and reply to almost none.
Read this cold email as that person:
"""
[subject line + email]
"""

Tell me honestly:
1) Would you reply? Yes, maybe or no, and why in one line.
2) Which sentence sounds generic or like a template?
3) Is there any claim you would not believe?
4) What one change would make you more likely to reply?
Do not rewrite the email. Just give me the feedback.

Fix what it flags yourself, then send. Writing the last edit by hand is what keeps the email in your voice. If you want ChatGPT to learn how you write in general, the voice setup in ChatGPT for client communication templates takes about ten minutes and works for cold emails too.

What one email looks like before and after

A made-up example, so you can see the difference the prompts make:

BEFORE (one-line prompt: "write a cold email to a dental clinic")
Subject: Grow your practice with our services!
I hope this finds you well. I came across your clinic and was impressed by your work. We help businesses like yours grow with innovative digital solutions. Would you be open to a 30 minute call next week to discuss how we can help?

AFTER (offer card + reason + rules)
Subject: your online booking form
Hi Sara, your new patient form asks for 9 fields before someone can pick a time. On a phone, that is a lot to fill in for a check-up. I build simple booking pages for small clinics, and I would cut it to 4. Want me to send a quick mockup of what that could look like?

Same sender, same clinic. The second one is shorter, specific, and asks for something easy to say yes to.

Before you send: the rules that keep you out of spam

A good email still fails if it lands in spam or breaks the rules. This is not legal advice, but these are the basics from the sources themselves:

• Say who you are. The FTC guide requires accurate From and Reply-To details, a valid physical postal address, and a clear way to opt out for commercial email in the US, and it makes no exception for business-to-business email. Opt-out requests must be honored within 10 business days.
• Never buy a list. Google’s sender guidelines say not to purchase email addresses and not to send messages to people who did not sign up. Cold email is a small, careful, hand-picked activity, not a blast.
• Start slow. The same Google guidelines recommend increasing sending volume gradually and avoiding sudden spikes. A handful of personal emails a day from your normal address is a very different thing from hundreds from a new domain.
• Check your local rules. If you email people in the EU or UK, the rules for unsolicited marketing can be stricter than in the US. Read your own regulator’s guidance before you start.

A simple line at the end works for most one-to-one emails: “If this is not relevant, just reply ‘no thanks’ and I will not follow up.”

Mistakes that waste the prompts

• Skipping the source text. If you do not paste real text about the prospect, ChatGPT invents a reason. That is how emails end up praising a “recent expansion” that never happened.
• Asking for a meeting in the first email. A 30 minute call is a big ask from a stranger. Offer something small first.
• Sending the first draft. The prospect check in Step 5 takes one minute and catches the line that sounds like a template.
• Writing to everyone. Ten emails to people with a clear reason beat a hundred to people without one.

Once someone replies, the real work starts. Turn the conversation into a clear offer with ChatGPT for proposal writing. And if inbound leads are piling up faster than you can answer them, an AI agent for lead qualification can sort them first, so your personal emails go to the best fits.

FAQ

Can I use these prompts on the free ChatGPT plan?

Yes. Every prompt here is plain text you paste in, with no file uploads or special features needed.

Do they work in Claude or Gemini?

Yes. They are plain instructions, so any chat assistant can follow them. Try the same prompt in two tools and keep the version that sounds most like you.

How many cold emails should I send per day?

There is no magic number. When every email needs a real reason, you will probably find that five to fifteen a day is what fits around your normal work, and a small, steady number also keeps your sending pattern calm.

Should I tell people I used AI?

You wrote the offer, chose the prospect, picked the reason and edited the final words. ChatGPT helped with the drafting, the way a spell checker helps with typos. What matters is that every fact in the email is true.

Write your first one today

Do Step 1 now, it takes ten minutes. Then pick one prospect you actually want to work with and run Steps 2 to 5. One good cold email you are proud of beats fifty you are not. Want more workflows like this, one practical win at a time? Join free at RobotsHelp.

Still building your prospect list? Our guide on how to use AI for lead generation shows how to find and research the people worth emailing first.

If you would rather hand outreach to someone else, compare the options in AI lead generation agency or DIY.

It is Monday morning, the CRM says you are on track, and you have a quiet feeling that half of those deals are not real. The fix is not a new tool. It is a short weekly review where ChatGPT does the sorting and you make the calls. Below are the ChatGPT prompts for sales pipeline management I would paste, in this order: check the export, find the stalled deals, give every deal a real next step, sanity-check the forecast, and write the follow-ups that move deals forward.

Imagine it is 9:30 on Monday, your coffee is still warm, and you already know the five deals that need you this week, the three you can close out, and the follow-up drafts are waiting in your outbox.

Want one practical AI workflow like this every week? Start free at RobotsHelp Launchpad.

What pipeline management with ChatGPT actually looks like

ChatGPT does not replace your CRM, and it should not update it for you. Its job in this routine is narrow: read a pipeline export, spot the gaps a tired human skims past, and draft the next messages. You stay the one who decides which deal gets pushed, parked or closed.

It works with an export from HubSpot, Pipedrive, Salesforce or a plain spreadsheet. OpenAI’s own guide to data analysis with ChatGPT says it can read CSV and Excel files and works best with clear column headers in the first row and one record per row. So one row per deal, plain column names, no merged cells.

Before you upload anything, strip what you do not need: personal phone numbers, private email addresses, anything your company rules keep inside the CRM. Deal name, stage, amount, dates and a short note are enough for every prompt below.

Step 1: Export the pipeline and let ChatGPT check the data

Export your open deals with these columns: Deal, Company, Stage, Amount, Close date, Last activity, Next step, Notes. Then start with a pure data check. No advice yet, just the facts that are broken.

You are my sales operations assistant. I am uploading my open pipeline as a CSV.
Columns: Deal, Company, Stage, Amount, Close date, Last activity, Next step, Notes.

Check the data only, no advice yet:
1) deals with a close date in the past
2) deals with an empty or vague next step ("follow up", "check in")
3) deals with no activity in the last 14 days
4) possible duplicate deals (same company, similar name)

Give one short table per check. Do not guess missing values. If a column is unclear, ask me one question before you start.

This first pass alone usually shows the problem. Close dates that already passed are the classic sign of a pipeline that looks healthier than it is.

Step 2: Test each deal against its stage

A stage name only means something if you know what has to be true to leave it. HubSpot’s sales pipeline walkthrough puts it well: each stage needs an objective exit criterion, such as “discovery completed with decision maker”, not “feels warm”. Write yours down once, then let ChatGPT check every deal against them.

Here are my pipeline stages and what must be true to leave each one:
- Discovery: we know the problem, the budget range and who decides.
- Proposal sent: the decision maker has seen the proposal.
- Negotiation: price and start date are being discussed in writing.

For each open deal, check whether the Notes show evidence that it meets the criteria of its CURRENT stage.
Mark each deal: Supported, Unclear or Not supported.
Quote the sentence from the notes that proves it. If there is none, say "no evidence".

Deals marked Not supported are not failures. They are just sitting one stage too high, and moving them back makes every number after this more honest.

Step 3: Sort the stalled deals into push, park or close

Most dead deals do not die with a no. They go quiet. This prompt forces a decision on every quiet one, so they stop inflating the total.

Take the deals with no activity in the last 14 days and sort them into three groups:

PUSH: clear buyer interest in the notes, but no next step. Say what is missing.
PARK: real, but the timing is later. Suggest a new close date and a check-in date.
CLOSE: no reply after two or more attempts, or the buyer said no or chose someone else.

One line per deal with the reason, based only on the notes. Do not invent buyer intent.

Step 4: Give every deal a real next step

The HubSpot walkthrough has a simple rule for pipeline reviews: if you cannot name the one scheduled next step for a deal, it does not belong in the forecast. “Follow up” is not a next step. “Client reviews the proposal with their CFO by Thursday” is.

For every deal marked Supported or Unclear, rewrite the Next step field as ONE buyer action with a date.
Good: "Anna reviews the proposal with her CFO by Thu 22 Oct."
Bad: "Follow up", "Send info", "Waiting".

If the notes do not show a buyer action, write "MISSING" and suggest the single question I should ask the buyer to get one.
Output a table: Deal, Old next step, New next step or MISSING, Question to ask.

Copy the new next steps into your CRM yourself. That two minutes of typing is also your sanity check that the suggestions make sense.

Step 5: Sanity-check the forecast

Your CRM probably shows a weighted number. In HubSpot’s default setup, each deal stage has a probability, and the weighted amount is the stage total multiplied by that probability. That is useful, but it trusts the stage. This prompt checks the evidence instead.

Take the deals with a close date in this month and split them into three groups:
COMMIT: decision maker named, price discussed, and a dated next step agreed by the buyer.
BEST CASE: two of those three.
PIPELINE: the rest.

Show the total amount per group. Then list every deal whose stage in my CRM looks more advanced than its group here, with the missing evidence.
Show your calculation so I can check it.

That last line matters. OpenAI recommends reviewing the code, outputs and assumptions whenever ChatGPT runs calculations on your data, and asking it to use a specific method when the result depends on one. If a total looks off, ask it to show the rows it summed.

Salesperson prompts for ChatGPT discovery call follow-up emails

The fastest way to move a deal out of discovery is a follow-up that lands the same day, in the buyer’s own words, with a clear next step. Paste your rough call notes straight in, typos and all.

Here are my notes from a discovery call: [paste notes].

Write a follow-up email to [name, role] that:
- thanks them in one line, no flattery
- sums up their problem in their own words (quote them where the notes allow)
- names the one priority I heard
- proposes the next step with two time options
- ends with one open question about who else should be involved

Under 150 words. Plain and warm, no hype words, no "I hope this finds you well".

For a deal that went quiet after the proposal, you need a different email. Many salespeople find that giving the buyer an easy way out gets more honest replies than another “just checking in”.

Write a short check-in for [name] about the [deal] proposal I sent on [date]. No reply since.
Three sentences max:
1) one specific detail from our last conversation
2) a simple yes or no question about the next step
3) an easy out: "If the timing changed, just reply 'later' and I will check back in [month]."
No guilt, no fake urgency.

This is for deals already in your pipeline. For cold outreach and multi-step sequences, see ChatGPT for sales outreach sequences.

Step 6: The Monday summary you send yourself

Finish the review with one short summary. Keep last week’s summary in the same chat, or paste it in, so ChatGPT can compare.

Summarize this pipeline review in 8 lines max:
- total open pipeline and the change since last week's summary (pasted below)
- deals moved forward, moved back, closed
- the 5 deals that need me this week, each with one action
- one risk I should not ignore

Plain text. Use only the data in this chat.

If you want that summary to arrive by itself every Monday, the n8n CRM automation workflows show how to build it as a small automation. And if your deal notes start life as messy call notes, summarizing meeting notes with ChatGPT gives you cleaner input for every prompt above.

What one deal looks like before and after

A made-up example, so you can see the shape of the output:

BEFORE
Deal: Northwind, onboarding package | Stage: Proposal sent | Amount: 8,400
Last activity: 19 days ago | Next step: follow up

AFTER THE REVIEW
Stage check: Unclear (no note that the decision maker saw the proposal)
Group: PUSH
Next step: MISSING. Ask: "Has the proposal reached your operations director yet, or should I walk them through it in 15 minutes?"
Forecast: PIPELINE, not COMMIT

Same deal, same CRM. The difference is that you now know what to do with it today.

Mistakes that make the review useless

• Letting ChatGPT guess. Every prompt above says “based only on the notes” for a reason. Without that, it fills gaps with plausible buyer intent.
• Reviewing every deal in detail. Let the prompts sort all of them, then spend your own time on the largest and the late-stage ones.
• Skipping the stage criteria. Step 2 is the one people drop, and it is the step that makes the forecast honest.
• Retyping the prompts every week. Save them once in a note or text snippet tool, and the whole review takes about the time of one coffee.

FAQ

Can I do this on the free ChatGPT plan?

For a small pipeline, yes: you can paste the table as text instead of uploading a file. File uploads and limits depend on your plan, so check OpenAI’s data analysis guide for what yours supports.

Do these prompts work in Claude or Gemini?

Yes. They are plain instructions with no ChatGPT-only features, so they work in any chat assistant that can read a table.

How often should I run a pipeline review?

Weekly, on the same day, works for most small teams and solo sellers. Daily turns into busywork, and monthly finds the stalled deals too late.

Is it safe to upload CRM data?

Follow your company’s rules first. Remove personal contact details you do not need, and check your ChatGPT workspace settings for how chats and files are stored.

Start with one review this week

Export the pipeline, run Step 1 and Step 3, and close out the deals that are already gone. That alone gives you a cleaner Monday. Want more workflows like this, one practical win at a time? Join free at RobotsHelp.

Need more deals in that pipeline? Start with a better first touch using our ChatGPT prompts for cold email outreach, short prompts that keep the first email personal.

The launch is in three weeks, the budget is set, and the brief says “fresh campaign idea”. You open a blank doc and type “campaign ideas for our product” into ChatGPT. Back come ten ideas that could belong to any company on earth.

Here is the short answer: the right ChatGPT prompts for marketing campaign ideas don’t ask for ideas first. They give ChatGPT your facts (product, audience, goal, budget, channels), ask for a few different angles, and then push the best angle into a one-page brief. ChatGPT gives you the raw material. You pick what fits the brand and what you can actually deliver.

Imagine Tuesday morning: by 10:30 you have three campaign directions, each with a headline, a hook and a first channel plan, and you walk into the planning meeting with something to react to instead of a blank page.

Want a free Launchpad lesson that turns this into a routine you reuse every campaign?

Why ChatGPT gives you generic campaign ideas

A model can only work with what you put in the chat. “Campaign ideas for our coffee brand” leaves out who buys, why now, what you can spend and where you will show it. So it fills the gaps with the average of everything it has read, and the average is exactly what you don’t want in a campaign.

OpenAI’s own prompt engineering guide comes back to the same few habits: clear instructions, the context the model needs, and a clear shape for the answer. For campaigns that context is a short campaign card. Write it once per campaign, paste it above every prompt below, and the answers get noticeably more specific.

Step 1: Write the campaign card (about 10 minutes)

Fill in the brackets with real facts only. If you don’t know a number yet, write “unknown” so ChatGPT doesn’t invent one.

CAMPAIGN CARD
Product or offer: [what it is, in one sentence]
Who it is for: [one clear audience, e.g. owners of small dental practices]
What they struggle with: [their problem in their own words]
Goal of this campaign: [e.g. 200 demo bookings, or 1,000 newsletter sign-ups]
Timing: [start date, end date, any season or event]
Budget: [amount, or "unknown"]
Channels we can actually use: [e.g. email list of 4,000, LinkedIn, one trade fair]
Brand voice: [three words, e.g. calm, practical, warm]
Never say: [claims we cannot make, competitors we never mention]
What worked before: [one past campaign and its result, or "none yet"]

Step 2: Ask for angles, not ideas

An angle is the reason someone should care right now. Ask for a few angles that are truly different from each other, so you get choices instead of ten versions of the same thought.

Using the campaign card above, give me 5 campaign angles that are clearly different from each other.
Use these 5 types, one each: a problem we solve, a before and after, a surprising fact from our own product, a season or moment, and a customer's point of view.
For each angle give: a working title (max 6 words), the one sentence the audience should remember, and why it fits this audience.
Do not invent numbers, results or customer quotes. If an angle needs a fact we don't have, say which fact.

That last line matters. It turns made-up statistics into a short to-do list of facts you can check.

Step 3: Turn the best angle into three concepts

I pick angle [number]. Develop it into 3 campaign concepts.
For each concept give: a headline, a hook for the first 3 seconds of a video, one visual idea we could shoot with a phone, and the call to action.
Keep each concept under 80 words. Write in our brand voice from the card.

Ask for a word limit every time. Short concepts are easier to compare in a meeting, and nobody falls in love with a paragraph they skimmed.

Step 4: Pressure-test it before the team sees it

Ideas sound best in the minute they are written. A quick test with a skeptical buyer saves you a round of polite nodding.

Act as a busy, skeptical person from our audience on the card.
Read concept [number] and tell me honestly:
1. Would you stop scrolling for this? Why or why not?
2. What is unclear after reading it once?
3. What would make you trust it more?
Then suggest 2 small changes that fix the weakest point.

Treat the answer as a fresh pair of eyes, not as market research. Real feedback still comes from real customers, so test the winner with a small budget or a part of your list first.

Step 5: Build the one-page campaign brief

Turn concept [number] into a one-page campaign brief with these headings:
Goal and how we measure it, Audience, Key message, Channels and what goes out on each, Timeline by week, What we need to produce, Open questions.
Use only facts from the card. Put anything you assumed under Open questions.

Now the idea has a plan around it, and the open questions show you exactly what to check with sales, legal or the client before anything goes live.

Prompts for campaign launch announcement emails

Most campaigns start with an email to the people who already know you. Once the brief is done, one more prompt gives you a first draft.

Write the launch email for this campaign to our existing list.
Give me 3 subject lines (max 45 characters), then one email of max 150 words: what is new, why it matters to the reader, one clear call to action.
Brand voice from the card. No exclamation marks, no claims that are not on the card.

Read it once out loud before it goes anywhere. If a sentence sounds like nobody at your company would say it, rewrite that sentence yourself.

Once the concept is chosen, the campaign still needs images. Our best AI tool for image generation guide shows which tool fits mood boards, ad visuals and social graphics, plus the rights checks to do before anything goes public.

Three mistakes that make the ideas worse

• Skipping the card. Without it, every prompt starts from zero and the answers drift back to generic.
• Asking for 50 ideas. Long lists feel productive but are hard to judge. Five different angles beat fifty similar ones.
• Publishing the first draft. ChatGPT can state things confidently that aren’t true, so every number, claim and name gets checked by a person before it goes out. OpenAI says this plainly in its help article on accuracy.

A simple weekly routine

For a recurring campaign (a monthly newsletter push, a quarterly offer) keep the card in a doc and update only what changed. Then run the same order every time: angles, concepts, pressure test, brief, launch email. It takes about half an hour, and the planning meeting starts with three real options on the table.

If you write social posts for the campaign too, the ChatGPT prompts for social media managers pick up where this brief ends. For the follow-up messages after a launch, see ChatGPT for sales outreach sequences, and if the campaign needs blog content, start with the ChatGPT prompts for an SEO content brief.

Not sure yet what your buyers actually care about? Start with how to use AI for market research, so the angles come from real evidence instead of guesses.

If the campaign comes with news worth telling reporters, here is how to use AI to write a press release from the same facts.

If the campaign needs proof from a real customer, see how to write a case study with AI without inventing a single result.

Turn it into a routine you keep

Learn the routine once and every new campaign starts from a plan instead of a blank page.

If the campaign needs video on a small budget, here is how AI UGC videos for small business work and how to start without faking customers.