tellaprompt

Prompts to summarize a document

Last updated

The failure mode of AI summaries isn't length — it's quiet invention: smoothed-over gaps, blended-in outside knowledge, confident statements the document never makes. The four templates below all counter it the same way: work only from the text, reference the location, and mark inference as inference.

Pick by purpose: layered (skim → study), decision-focused (what do I need to DO?), pinpoint-referenced (when you'll be quoted), or comparative (many documents, one matrix).

The four templates

Layered summary: 3 sentences, 3 paragraphs, full notes

You don't know yet how deep you need to go — so buy all three depths at once.

Summarize the document below in three layers:
1) **Three sentences** — what it is, what it claims/decides, why it matters.
2) **Three paragraphs** — the argument or content in order, main evidence included.
3) **Structured notes** — per section: heading, core statements as bullets, notable numbers/dates/names.

Rules: only what's in the text; where the document is vague, say "the document leaves open…" instead of resolving it; keep the author's key terms (don't paraphrase defined terms).

Document:
[PASTE — FOR LONG DOCS IN CHUNKS: "PART 1/3" ETC., SUMMARY AFTER THE LAST PART]

Why this works: Three depths in one pass means the skim can escalate to study without re-prompting; leaves-open phrasing preserves the document's actual uncertainty.

Output formatContext firstFact density

Decision-focused extraction

A 60-page document, one question: what does this mean for what we do next?

Read the document below through one lens: [MY DECISION/SITUATION, E.G. "DO WE RENEW THIS CONTRACT?" / "DOES THIS REGULATION AFFECT OUR PRODUCT?"].

Deliver: 1) directly relevant passages — quoted, with location, and one line each on WHY relevant; 2) obligations/deadlines/conditions as a table (what, who, by when, consequence); 3) what the document does NOT regulate or answer that my decision needs — listed explicitly; 4) verdict template: "Based on this document alone: [X]. Not covered and still needed: [Y]."
Ignore everything not relevant to the lens — completeness is not the goal, relevance is.

Document:
[PASTE]

Why this works: The single-lens rule flips summarization from compression to retrieval; the not-covered list is what stops "the document doesn't forbid it" misreadings.

Context firstOutput formatAsk-first loop

Faithful summary with pinpoint references

You'll cite this summary in a memo or meeting — every sentence needs a way home.

Summarize the document below with references. Length: [N] words.

Rules: every substantive statement ends with its location (§/page/paragraph); direct claims quoted where wording matters; anything you conclude that the text doesn't literally say is marked [INFERENCE]; anything ambiguous gets both readings, not your pick. End with: 3 statements a hasty reader would wrongly take from this document (the likely misreadings).

Document (numbered sections/pages if possible):
[PASTE]

Why this works: Pinpoint-or-silence makes the summary verifiable in seconds; the likely-misreadings section is cheap insurance for exactly the meeting where you'll be challenged.

Fact densityOutput formatBrutal critic

Multi-document comparison matrix

Five vendor offers, three studies, four policy drafts — one comparable picture.

Compare the [N] documents below on the dimensions that matter to me: [LIST DIMENSIONS, E.G. PRICE, TERMINATION, LIABILITY / METHOD, SAMPLE, FINDING].

Deliver: 1) matrix — dimension × document, cell content quoted or tightly paraphrased with location, empty cells marked "not addressed" (never filled by analogy); 2) per dimension: one line on the biggest spread and where the outlier is; 3) traps: places where documents use the same word differently or different words for the same thing; 4) what I'd still need to ask each document's owner.

Documents:
[PASTE, CLEARLY LABELED: DOC A = …, DOC B = …]

Why this works: Not-addressed cells and the same-word-different-meaning trap list are where comparison matrices usually lie; making both explicit keeps the matrix honest.

Output formatFact densityContext first

Frequently asked questions

How do I summarize documents longer than the chat allows?

Chunk them with labels ("Part 1/4"), have the model acknowledge each part with a two-line running digest, and request the full summary after the last chunk. For very long or many documents, tools with file upload and large context windows do this natively — the prompt rules transfer unchanged.

Can I trust an AI summary of a contract or legal text?

Trust it as a map, not as advice: the pinpoint-reference template makes every statement checkable against the clause, which is the point. Legal significance stays with a professional — see the lawyer page for review-grade templates.

Why do summaries sometimes contain things not in the document?

Models blend document content with background knowledge unless explicitly fenced — that's why every template here says "only from the text" and marks [INFERENCE]. The rules aren't decoration; without them, blending is the default behavior.

What summary length should I ask for?

Decide by use: 3 sentences to triage, ~15 % of original for working summaries, pinpoint-referenced at any length for citable ones. The layered template sidesteps the choice by giving three depths at once.