PromptForge/Canon/Custom AI prompts
Canonical prompt · verified · 2026-07-18

The canonical prompt for summarizing long documents

A summarization prompt that preserves decisions, numbers, and disagreements instead of flattening everything into vague bullet points — with a claims-check that prevents summary hallucination.

What this prompt produces

A reusable prompt that turns any long document (report, transcript, paper, thread) into: a 3-sentence executive summary, decisions/conclusions with their reasoning, every number that matters with its context, open questions and disagreements preserved, and action items with owners — all traceable to the source, nothing invented.

The canonical prompt

You are a chief-of-staff-grade analyst. Executives act on your summaries without reading the source, so a wrong or missing fact in your summary becomes a wrong decision.

Summarize this document:
[PASTE DOCUMENT]
Reader: [who will read this and what they'll do with it]

OUTPUT FORMAT:
## TL;DR (3 sentences max)
The single most important takeaway, the second, and what happens next.
## Decisions & conclusions
Each decision WITH its stated reasoning ("chose X because Y"). If reasoning is absent in the source, write "(reasoning not stated)".
## Numbers that matter
Every consequential figure with context: "[metric]: [value] ([comparison/period])". Skip decorative numbers.
## Disagreements & open questions
Preserve who disagreed about what. Unresolved items listed as questions, not smoothed into false consensus.
## Action items
[owner] → [action] → [deadline if stated]. Only items the source actually assigns.

HARD RULES:
- Every claim in your summary must be traceable to the source. NOTHING may be inferred, averaged, or "reasonably assumed".
- Preserve hedges: if the source says "probably", your summary says "probably". Removing uncertainty is falsification.
- Disagreements are content, not noise — never merge opposing views into a neutral statement.
- If a section has no content in the source, write "None in source" — do not fill it.

SELF-CHECK BEFORE OUTPUT: pick your 3 most load-bearing claims and verify each exists in the source text verbatim or near-verbatim. If any doesn't, delete it.
This prompt is generic — on purpose.

Copy it freely. If you want a version tailored to YOUR project (name, industry, features, style), generate one in 20 seconds:

Forge it for my project →

Why this prompt works

  • Naming the reader and their next action focuses selection — summaries fail by summarizing for nobody.
  • 'Preserve hedges' and 'disagreements are content' target the two ways summaries silently lie.
  • The traceable-claims rule plus the 3-claim self-check measurably cuts summary hallucination.
  • Structured sections turn 'summarize' from a compression task into an extraction task — models are far better at extraction.

What weaker prompts get wrong

  • 'Summarize this' with no structure — returns generic bullets that flatten decisions into topics.
  • Letting the model resolve disagreements — false consensus is the most dangerous summary failure.
  • No hedge preservation — 'might work' becomes 'works' after two summary hops.

Related canonical prompts