Summaries that keep what mattered, and rewrites that change form without changing meaning. Below are 7 copy-ready prompts. Fill in the [BRACKETS], copy, and paste into ChatGPT, Claude, Gemini or any capable assistant.
A bad summary is not shorter, it is vaguer. Models default to vaguer because generality is safer than choosing what matters.
The 7 prompts
Summarise a long document at three depths
Get a summary sized to how much the reader needs.
Summarise this at three depths. DOCUMENT: """ [PASTE] """ WHAT IT IS: [REPORT / PAPER / CONTRACT / TRANSCRIPT / ARTICLE] WHO THE SUMMARY IS FOR: [ROLE] WHAT THEY NEED IT FOR: [THE DECISION OR TASK] Produce: **30 SECONDS** (max 60 words): what it says and what the reader should do about it. **3 MINUTES** (max 350 words): main claims, key evidence, conclusions, and anything that affects this reader specifically. **FULL** (as long as needed): section by section, preserving structure, with the reasoning not just the conclusions. Then: - WHAT THE DOCUMENT DOES NOT SAY that the reader might assume it does - THE PART MOST RELEVANT to their stated task, quoted directly - ANYTHING SURPRISING or that contradicts what they would expect - CONFIDENCE AND HEDGING - where the document is tentative versus definite. A summary that flattens 'the evidence weakly suggests' into 'X causes Y' is worse than no summary. Rules: - Do not add interpretation. If the document does not conclude something, do not conclude it. - Preserve numbers exactly, including units and time periods. - Where the document contradicts itself, say so rather than picking one. - Quote directly for anything a decision will rest on.
What you get: Three depths of summary, plus what the document does not say and where its hedging matters.
Tip: Preserving hedging is what makes a summary trustworthy. Summaries that turn 'may be associated with' into 'causes' are how misinformation starts inside organisations.
Rewrite something at a different length
Expand or compress a piece to a target length properly.
Rewrite this at [TARGET LENGTH]. Current length is [CURRENT LENGTH]. TEXT: """ [PASTE] """ WHAT MUST SURVIVE: [ESSENTIAL POINTS] READER: [WHO] PURPOSE: [WHAT IT NEEDS TO ACHIEVE] IF COMPRESSING: - Cut whole ideas before trimming sentences. Say which ideas you cut. - Keep the specifics - numbers, names, examples - and cut the connective prose. Specificity is what makes short writing land. - Do not compress by generalising. 'Three teams reported the same issue' should not become 'several stakeholders raised concerns'. - Where an idea cannot survive compression without becoming misleading, keep it in full and cut something else. Flag this. IF EXPANDING: - Do not pad. Add only: examples, evidence, addressed objections, implications, and context the reader needs. - Never expand by restating. If you cannot add substance, say so and deliver it shorter than asked. - Mark [NEED FROM YOU] where expansion requires information I have not given. Do not invent examples. Either way, output: - The rewritten version at target length - WHAT CHANGED - cut or added, as a list - WHAT WAS LOST OR UNVERIFIED - THE HONEST NOTE - if this text does not want to be the target length, say so. Some arguments cannot be made in 200 words, and a 2,000-word version of a 400-word idea is padding wearing a suit.
What you get: A rewrite at target length with a change log and an honest note if the target length is wrong for the content.
Tip: The refusal to pad is the useful part. Being told 'this is a 400-word idea' saves you publishing 2,000 words of filler.
Summarise a long meeting or interview transcript
Extract what matters from hours of talking.
Summarise this transcript. TRANSCRIPT: """ [PASTE] """ WHAT IT IS: [MEETING / INTERVIEW / USER RESEARCH / PANEL / CALL] PARTICIPANTS: [NAMES AND ROLES IF KNOWN] WHAT I NEED FROM IT: [THE PURPOSE] Produce: 1. THE SHORT VERSION - 100 words. What happened and what matters. 2. BY TOPIC not chronologically - people circle back to subjects, and a chronological summary scatters each topic across the document. Group everything said about each topic together. 3. WHO SAID WHAT - for anything contested or consequential, attribute it. Quote directly where the exact words matter. 4. AGREEMENT AND DISAGREEMENT - where participants agreed, where they did not, and where a disagreement was left unresolved. Unresolved disagreements are the most important thing in most transcripts and are usually smoothed over in summaries. 5. QUOTABLE MOMENTS - the strongest verbatim lines, with attribution. Useful for research, marketing and reports. 6. WHAT WAS AVOIDED - topics raised and deflected, questions not answered, subjects nobody brought up that you would expect. Be careful here; mark as observation, not accusation. 7. ACTIONS AND COMMITMENTS - anything anyone said they would do. Flag as UNASSIGNED where no owner was named. 8. IF THIS IS RESEARCH - the patterns across participants, and the outliers. Distinguish what one person said from what several said; a single vivid quote is not a pattern. Rules: - Do not clean up quotes into something people did not say. Trim with ellipses; do not rewrite. - Do not resolve a disagreement the participants left open. - Mark [INAUDIBLE] or [UNCLEAR] where the transcript is garbled rather than guessing.
What you get: A topic-grouped summary with attributed quotes, unresolved disagreements preserved, and single opinions distinguished from patterns.
Tip: Grouping by topic rather than time is the single biggest improvement to transcript summaries. Conversations wander; summaries should not.
Compare several documents and find the differences
Work out where multiple sources agree and conflict.
Compare these documents. DOCUMENTS: """ [PASTE EACH, SEPARATED BY --- WITH A LABEL] """ WHAT THEY ARE: [PROPOSALS / VERSIONS / VENDOR RESPONSES / RESEARCH PAPERS / POLICIES] WHAT I NEED TO DECIDE OR UNDERSTAND: [PURPOSE] Produce: 1. COMPARISON TABLE - the dimensions that matter for my purpose as rows, each document as a column. Derive the dimensions from the documents and my purpose, not from a generic template. Mark [NOT ADDRESSED] where a document is silent - silence is a finding. 2. WHERE THEY AGREE - substantively, not just superficially. Note where they agree on a conclusion but for different reasons; that matters. 3. WHERE THEY CONFLICT - direct contradictions. For each: what each says, quoted, and what would resolve it. Do not average conflicting claims into a middle position. 4. WHERE THEY TALK PAST EACH OTHER - the subtler and more common case. Different definitions of the same term, different scope, different assumptions, different time periods. These look like disagreement but are not, or look like agreement but are not. 5. WHAT ONE COVERS THAT OTHERS DO NOT - gaps and unique contributions. 6. RELIABILITY DIFFERENCES - where the documents differ in evidence quality, recency, scope, or apparent interest in the outcome. Note any document with an obvious stake in its own conclusion. 7. FOR MY PURPOSE - the three differences that most affect my stated decision. 8. WHAT NONE OF THEM ADDRESSES - the question I would still need answered. Quote directly for every conflict. Do not paraphrase a disagreement.
What you get: A purpose-driven comparison table with quoted conflicts, talking-past-each-other cases separated out, and the gap none of them fills.
Tip: Point 4 is where most comparisons go wrong. Two documents using 'active user' differently look like they disagree about the numbers when they are measuring different things.
Extract structured data from unstructured text
Turn prose into a table you can work with.
Extract structured data from this text. TEXT: """ [PASTE] """ WHAT I WANT EXTRACTED: [THE FIELDS] OUTPUT FORMAT: [table / CSV / JSON] WHAT TO DO WITH MISSING VALUES: [leave blank / write NOT STATED / omit the row] Rules for extraction: - Extract only what is stated. Never infer a value from context and present it as extracted. If something is implied but not stated, put it in a separate 'inferred' column marked clearly. - Preserve the original wording for anything categorical. Do not normalise 'quite large' into 'large' unless I gave you a category list. - Preserve numbers exactly as written, with their units. Do not convert units. - Where the text gives a range, keep the range. Do not take a midpoint. - Where the same fact appears twice with different values, include both and flag the conflict. - Where a value is ambiguous, extract it and add a note column explaining the ambiguity. Output: 1. THE STRUCTURED DATA in my requested format. 2. EXTRACTION NOTES - a row-by-row note for anything that was not a clean extraction: ambiguous, conflicting, inferred, or partially stated. 3. WHAT I COULD NOT FIND - fields I asked for that the text does not contain, and whether the text suggests they exist elsewhere. 4. CONFIDENCE - overall, and specifically which fields were reliably extractable and which required judgement. 5. SOURCE QUOTES - for any field where accuracy matters, the phrase you extracted it from, so I can verify without rereading. Accuracy over completeness. A blank cell is better than a plausible guess.
What you get: Structured output with inferences separated from extractions, conflicts flagged, and source quotes for verification.
Tip: Keeping inference in its own column is the rule that makes extracted data trustworthy. Once a guess is in the same column as a fact, nobody can tell them apart later.
Write an abstract or executive summary
Produce the summary that goes at the top of a long document.
Write an executive summary for this document. DOCUMENT: """ [PASTE] """ WHO READS THE SUMMARY: [ROLE AND WHAT THEY DECIDE] WILL THEY READ THE FULL DOCUMENT: [probably not / skim it / yes] LENGTH: [WORDS] Write it so that it stands alone. Assume most readers will read only this. Structure: 1. THE CONCLUSION FIRST - what this document concludes and what should happen as a result. Not what the document is about; what it found. 2. WHY IT MATTERS - the consequence of acting or not acting, in the reader's terms. 3. THE BASIS - the two or three findings the conclusion rests on, each with its supporting figure. 4. WHAT IS UNCERTAIN - the limitations that would change the conclusion. Include this even when it weakens the case; a summary that hides its caveats loses credibility when someone reads the full document. 5. THE ASK - the specific decision, action or approval needed, and by when. Rules: - Never open with 'This document...' or 'The purpose of this report is...' - Do not describe the document's structure. Nobody needs a summary of the table of contents. - Every number must appear in the document. Do not compute new ones. - It must be readable by someone with no prior context. - No jargon undefined in the summary itself. Then tell me: the one sentence the reader will remember, and whether it is the right one.
What you get: A standalone conclusion-first summary with the basis, the caveats and the ask, plus the sentence the reader will retain.
Tip: 'Do not describe the document's structure' removes the most common failure. An executive summary that says what each section covers has summarised nothing.
Turn a long piece into social and email versions
Repurpose one article into several formats without it reading as recycled.
Repurpose this piece. ORIGINAL: """ [PASTE] """ FORMATS I WANT: [LinkedIn post / X thread / newsletter section / short video script / Instagram caption / other] MY AUDIENCE ON EACH: [WHO FOLLOWS ME WHERE] MY TONE: [DESCRIPTION, OR PASTE A SAMPLE POST] First identify: - THE ONE IDEA - the single most interesting thing in the piece. Repurposing works by extracting one idea properly, not by compressing all of them badly. - THE SECOND AND THIRD IDEAS - each could be its own post later. List them so I get several weeks out of one article. Then for each requested format: - The post, written natively for that platform, not adapted from the others - Why this format suits this idea, or why it does not - The opening line, which is what determines whether anyone reads the rest - Whether and how to link to the full piece Format-specific rules: - LinkedIn: the first two lines show before 'see more'. Put the whole hook there. No engagement bait. - X thread: each post must stand alone and earn the next. Number them only if the order is essential. - Newsletter: assume a warmer, more personal register than public social. - Video script: written to be spoken. Short sentences, no clauses that only work on the page. Rules: - Do not post the same text to every platform. - No 'I wrote a new blog post' as an opening. - Each version must deliver value on its own, even to someone who never clicks through. - No hashtag walls, no 'thoughts?', no fake vulnerability opening.
What you get: One idea extracted properly per platform, natively written, plus two more ideas banked for future posts.
Tip: Extracting one idea rather than compressing the whole article is the difference between a post worth reading and an ad for a post.
Where AI actually helps here
- Compressing long documents when you say what you need out of them
- Rewriting for a different audience or reading level
- Extracting a specific thing — every commitment, every date, every objection
Where it falls down
- Deciding what matters without being told. Its priorities are not yours
- Long documents in one pass. Quality drops well before the context limit
- Preserving hedges. “Results may suggest” becomes “results show” with alarming regularity
The mistake almost everyone makes: Not saying what the summary is for
“Summarise this” produces a proportional précis. “Summarise this for someone deciding whether to renew the contract” produces the four things that bear on renewal. Same document, entirely different output, one extra clause.
Free tool: Prompt Improver
Runs in your browser. No sign-up, nothing uploaded.
Questions people ask
How do I summarise a document too long for the context window?
Chunk it, summarise each chunk with the same instruction, then summarise the summaries. Tell it which chunk it is seeing and not to conclude until the end, or it writes a conclusion for part two of five.
Why do AI summaries lose the important bit?
Because importance is relative to a purpose you did not state. It defaults to proportional coverage — the longest section becomes the biggest paragraph, whether or not it mattered.