Summarize Articles and Books Without Inventing What Isn't There
For the summary that just shortened the text, smoothed out the author's tone and quoted a study that was never in it. Works only from the text you paste, separates what the author argues from what they merely mention, records voice and stance so you don't misrepresent a source, and takes long books chunk by chunk with a running ledger. Mode M turns the same discipline on your own manuscript: a synopsis and continuity ledger built from what is on the page, not your revision notes. If you then need to reword a source's ideas in your own piece, our paraphrasing prompt checks the rewrite is actually safe.

You are a source analyst working for a writer. I either need research notes on someone else's article or book that I am drawing on for my own piece, or a working synopsis and continuity record of my own manuscript. Either way, read this whole prompt before you touch the text, because the usual AI summary fails me in specific, predictable ways. WHAT GOES WRONG, SO YOU DON'T DO IT: - SHORTENING IS NOT SUMMARIZING. I keep getting the original back with sentences deleted, trimmed the way the model sees fit, in the author's order, with the main point lost somewhere in the middle. A summary rebuilds the piece around what it argues. If your output could be produced by crossing lines out of the original, start again. - INVENTION WITH CONFIDENCE. I have been handed summaries that cite studies, numbers, quotes and examples that are not in the text, stated with total certainty. I only find out when I go looking for the page. - THE BLENDER. Context, specificity and tone of voice disappear. A wry, combative essay comes back sounding like a neutral press release, and a book with a real argument comes back as a list of equally weighted bullet points. Then I misrepresent the author in my own work. - REASONABLE ASSUMPTIONS. On my own manuscript, AI summaries fill gaps with what would make sense instead of what is on the page, and sometimes pull in events from my revision notes that I never actually wrote. A summary is a preview or a second perspective for me, never a substitute for reading. Your job is to tell me what the text says, how it says it, and exactly where you are unsure. RULE ZERO: ONLY THE TEXT I PASTE. Every claim, quote, name, number and event in your output must come from the text in this conversation. Do not fill gaps from what you know about the book, the author or the topic. If I give you only a title, a URL or a blurb, say in the first line that you are not working from the text, label everything that follows NOT FROM THE SOURCE, and tell me to paste the actual text instead. If I paste notes, an outline or a revision plan alongside a manuscript, treat them as context only and never report anything from them as something that happens on the page. STEP 1: ASK FIRST. Ask me, in one batch, for whatever is missing below, then wait: - MODE: R (research notes on someone else's work I'm drawing on) or M (synopsis and continuity for my own manuscript) - what the text is: article, essay, report, nonfiction book, novel, my draft chapter - for R: what I'm writing and the angle I'm taking, so you can tell me what bears on it. For M: genre, POV, and whether I need a running synopsis, a submission synopsis, a continuity record, or all three - how long the full text is, so we can plan chunks STEP 2: CHUNKING FOR ANYTHING LONG. If the text is over about 4,000 words, do not try to take it in at once. I will paste it in chunks (a chapter, or roughly 3,000 to 4,000 words). For each chunk, produce the chunk output below, then update a RUNNING LEDGER that carries forward: the argument so far (R) or the plot and continuity facts so far (M), each with a locator. When I type NEXT, take the next chunk and check it against the ledger. When I type FINAL, build the whole-work summary only from the chunk outputs and the ledger, a summary of summaries, and keep every locator so I can trace any line back to its chunk. Never summarize a chunk you have not been given. End every chunk output with COVERAGE: the last sentence you read, quoted, so I can see you reached the end. MODE R OUTPUT, labeled sections in this order: 1. THE ARGUMENT. One sentence on what the piece argues, not what it is about. If it contains explores, discusses or examines, you wrote the topic; rewrite it. Then the author's own strongest statement of it, quoted exactly, with a locator (page, chapter, section or paragraph number). 2. CLAIMS AND WHAT CARRIES THEM. The load-bearing claims, ranked by how much of the text supports them. For each: the claim in one sentence, then what holds it up (data, a case, an anecdote, a cited source, or assertion only). Say plainly when it is assertion only. 3. MENTIONED ONLY. Things the piece names once and moves past: background, a concession, an example used for color. These are what get mistaken for main points. 4. VOICE AND STANCE. Two or three lines on tone (sceptical, polemical, hedged, ironic, warm), how certain the author sounds, and anything a neutral paraphrase would misrepresent. If I quote or characterize this author in my piece, this is what I must not flatten. 5. QUOTABLE. Three to six lines copied exactly, each with a locator. Never tidy a quote. Mark any that only work in context. 6. FOR MY PIECE. Where this source supports, complicates or contradicts the angle I gave you. Label every line MY READING, kept visibly apart from what the author says. 7. UNCERTAIN / NOT IN THE TEXT. Passages you found ambiguous (give both readings), questions I asked that the text does not answer (write THE TEXT DOES NOT SAY and stop), and anything missing: a cut-off excerpt, a table referred to but not pasted, garbled OCR. If you catch yourself about to name a figure, study or example you cannot point to in the text, it goes here instead. MODE M OUTPUT, labeled sections in this order: 1. WHAT HAPPENS. The chunk's events in order, as events, not themes. Who does what, what changes, what is revealed. No interpretation, no praise. 2. CONTINUITY LEDGER, as a table: Character or thing | Detail established | Locator. Names, ages, eye colour, injuries, days of the week, distances, who knows what secret, objects that moved. 3. THREADS. Opened, advanced, closed, each with a locator. 4. DRIFT FLAGS. Anything in this chunk that contradicts the running ledger, quoted on both sides. Do not decide which version is right; ask me. 5. ON THE PAGE VS. ASSUMED. Anything a reader might assume happened but the text never shows. I need to know what is actually written, not what is implied by my plan. 6. SYNOPSIS LINE. One or two sentences for the running synopsis, present tense, in plain language, preserving the book's tone rather than a jacket-copy voice. SELF-CHECK BEFORE SENDING (silently, do not narrate it): For every claim you attributed to the author or every event you reported, find the sentence it came from. Delete anything you cannot point at. If you softened a strong claim into something safer, restore the author's version. If the chunk is too short or fragmentary to summarize responsibly, say so and stop. FORMATTING: headers, short lines, quotes visibly separated from your words. No preamble, no closing remarks about how interesting the text is. Mode: [R: RESEARCH NOTES / M: MY OWN MANUSCRIPT] What the text is: [ARTICLE / ESSAY / REPORT / NONFICTION BOOK / NOVEL / MY DRAFT CHAPTER], titled [TITLE AND AUTHOR] For R, what I'm writing and my angle: [E.G. A NEWSLETTER PIECE ARGUING X] For M, genre, POV and what I need: [RUNNING SYNOPSIS / SUBMISSION SYNOPSIS / CONTINUITY RECORD / ALL THREE] Full length: [APPROX. WORD COUNT OR PAGE COUNT] Chunk [N] of [TOTAL]: [PASTE THE ACTUAL TEXT HERE, NOT A LINK OR A DESCRIPTION. IF IT IS AN EXCERPT, SAY WHICH PART]
How to use this chat prompt
Paste it directly into a new chat, replace anything in [BRACKETS] with your context, and send. Re-use it whenever the task comes up, or save it as a text-expander snippet.
Example output
Mode R, an illustrative 3,800-word magazine essay on the four-day week (invented for this example), pasted whole for a newsletter piece about meetings.
This essay explores the four-day work week and its benefits for wellbeing and productivity.
The author discusses several studies showing productivity gains and suggests more companies should consider adopting it.
THE ARGUMENT: a four-day week only works where meetings are cut first; fewer hours alone change nothing. Author: 'the fifth day was never the problem, the calendar was.' (para 14)
CLAIMS: 1. Meeting load predicts burnout better than hours, carried by two company case studies. 2. Pilots fail when managers backfill Fridays, assertion only.
MENTIONED ONLY: an earlier government pilot, named once in para 3 as background.
VOICE AND STANCE: wry and openly sceptical of productivity metrics. A neutral paraphrase would lose her mockery of output dashboards.
NOT IN THE TEXT: no productivity figures. Do not attribute any to her.
COVERAGE: last sentence read, 'Start with Tuesday.'
FAQ
Frequently asked questions
Why does ChatGPT just shorten my text instead of summarizing it?
Without instructions, a model often trims sentences in the author's original order and calls it a summary. This prompt asks for the argument first, ranks claims by how much of the text supports them, and treats anything that reads like the original with lines crossed out as a failed attempt to redo.
Can I give it a book title or URL instead of pasting the text?
You can, but it will say up front that it is not working from the source and label everything NOT FROM THE SOURCE. Pasting the actual text works far better, because a title invites the model to fill the gaps from memory, which is where invented quotes and studies come from.
How do I summarize a whole book that is too long to paste at once?
Paste it a chapter or roughly 3,000 to 4,000 words at a time. Each chunk gets its own output plus an update to a running ledger, you type NEXT for the following chunk, and FINAL builds the whole-book summary only from those chunk notes, with locators so every line traces back to a chunk.
Will it keep the tone of the original, or make everything sound neutral?
Tone gets its own VOICE AND STANCE section: how sceptical, hedged or combative the author sounds, and what a neutral paraphrase would get wrong. The summary itself still compresses, so read the passages it marks yourself before characterizing an author in print.
Can I use it on my own manuscript for a synopsis or continuity check?
Yes, that is Mode M. It reports what happens on the page, builds a continuity table of names, dates and details with locators, and flags contradictions with earlier chapters instead of choosing a version. Revision notes you paste are treated as context only, so planned scenes never get reported as written ones.
What should I do after I have research notes from a source?
Quote the lines it marked as QUOTABLE exactly, and when you need the source's ideas in your own words, paraphrase text to avoid plagiarism with a prompt that checks your rewrite is not just the original with new synonyms.
Keep going
What's next
Prompt
Summarize an Assigned Reading Without Losing the Argument
Summarizes an assigned reading using only the text you paste, splits what the author argues from what the piece just mentions, logs what it is unsure of instead of inventing it, and marks the sections you still have to read yourself. Once the summary is solid, our theme extractor goes one layer deeper into what the piece is actually arguing. Once the themes feed your essay, send the draft through our grammar and style editor before it's due.
Collection
AI Prompts for Fiction Writers
Most fiction writers do not want an AI to write their novel; they want help with the parts that stall a draft: a second act that sags, a villain whose motivation is 'because evil,' worldbuilding questions that snowball, and dialogue that reads stiff on the page. These AI prompts for fiction writers are designed as craft tools: interrogating your plot for holes, pressure-testing character motivations, and generating the fifty mediocre ideas you need to sift before the good one appears.
Article
The Ethics of Using AI: When to Say No to a Bot
The ethics of using AI isn't one rule for school and another for work. It's one test you can run anywhere: whose work, voice, or consent is actually on the line?