Most summaries drop the one detail that made the original worth reading. The fix is not a longer prompt; it is a more specific one. A good summary prompt names the reader, a keep-list (facts that must survive), and a drop-list (what is safe to lose). Below are 12 copy-ready templates plus a quick guide to which model to point them at in June 2026.
TL;DR
- A reliable summary prompt carries six parts: audience, goal, voice, constraints, format, and 1-2 tone examples.
- Paste the actual text, not just a link, when you can — pasted transcripts and PDFs summarize more faithfully than URLs the model has to fetch and guess at.
- For one short article, any chat tier works. For a 30-page report or a 45-minute talk transcript, you want a 1M-token model: Claude Opus 4.7 / Sonnet 4.6, Gemini 3.1 Pro, or GPT-5.5 (full 1M only on the $200 ChatGPT Pro tier).
- Always run a fidelity check (template 11) before you publish. The models are confident, not always correct.
Which model for which summary (June 2026)
In-app context decides how much you can paste at once. These are the limits as of June 2026; verify on the vendor’s page before relying on a number.
| Source you’re compressing | Best fit | Why |
|---|---|---|
| One article or email | ChatGPT Free/Plus, Claude Free/Pro, Gemini free | Fits easily; speed matters more than window size |
| 30-page report or whitepaper | Claude Sonnet 4.6 (Pro $20) or Gemini 3.1 Pro (AI Pro $19.99) | 1M-token context standard; holds the whole doc |
| 45-min talk / podcast transcript | Gemini 3.1 Pro | Paste the YouTube transcript text, not the link, for a faithful read |
| Conflicting multi-source synthesis | Claude Opus 4.7 (Max) | Best at flagging contradictions and gaps, not just compressing |
| Long doc inside ChatGPT | ChatGPT Plus handles ~320 pages in-app; full 1M context needs the $200 Pro tier | Plus truncates very long pastes silently |
A note on names and tiers as of June 2026: Google AI Pro ($19.99) is the renamed “Gemini Advanced” and now ships Gemini 3.1 Pro with a 1M-token window. Claude Pro ($20) runs Sonnet 4.6; Opus 4.7 needs Max ($100+). ChatGPT Plus ($20) defaults to GPT-5.5, but the in-app window holds roughly 320 pages — the full 1M-token context only opens on the $200 Pro plan.
Who this is for
Newsletter writers, marketers compressing whitepapers, executives skimming research, and students pulling the essence out of papers.
When not to use these prompts
Skip them for legal, medical, or safety content where every word carries liability. And do not use them to manufacture false balance — a “both sides” summary of a one-sided source is a distortion, not a compression.
Prompt anatomy: the six parts that matter
Every summary prompt should carry six elements. Drop one and the output drifts toward generic.
- Audience: one specific reader, not “general public.”
- Goal: one outcome — read, click, agree, or share.
- Voice: 2-3 anchor adjectives.
- Constraints: word count, banned phrases, must-include facts.
- Format: paragraph, bulleted, headed, or table.
- Examples: 1-2 tone samples. This is the strongest lever for matching voice; “be friendly” is noise by comparison.
12 copy-ready prompt templates
In each template, replace any [bracketed] placeholder with your own value. Paste the source text directly above or below the prompt.
1. Length-bound summary
Summarise this in exactly [number] words. Audience: [one specific reader]. Must include: (a) main claim, (b) one supporting fact, (c) one nuance. Drop: history, methodology, hedges. The output word count must match the target.Swap: the word count and the reader.
2. Paper to layperson
This is a research paper. Summarise in 150 words for a non-specialist: (1) Question asked, (2) What they did, (3) What they found, (4) One caveat, (5) Why anyone should care. No "the authors argue" framing — state direct claims.3. Talk to tweet thread
Convert this 45-min talk into an 8-tweet thread. Tweet 1: hook. Tweets 2-6: core arguments, one per tweet. Tweet 7: counter-point or caveat. Tweet 8: where to watch or read more. Each tweet 280 characters or fewer.Paste the transcript, not the video link — pasted text summarizes far more faithfully than a URL the model has to fetch.
4. Email to executive TL;DR
Above this 1000-word email, add a 3-line TL;DR: (1) Decision needed or FYI, (2) Recommendation if any, (3) Deadline. Keep the original body below intact.5. Report to bullet brief
Summarise this 30-page report as: (1) 5 bullet findings, (2) 3 bullet recommendations, (3) 1 bullet of "what surprised us". Each bullet 25 words or fewer. Drop methodology unless it changes the interpretation.6. Multi-source synthesis
Read these 3 sources on the same topic. Produce a 200-word summary that: (a) names points of agreement, (b) names points of disagreement, (c) flags where sources contradict each other, (d) names the gap none of them cover.This is the one to run on Claude Opus 4.7 — it is the strongest at flagging contradictions rather than averaging them away.
7. Skeptic’s summary
Summarise this piece as a skeptical reader would. Compress to 150 words, leading with the strongest counter-evidence. The author would not love this summary — that is the point.8. Comparison summary
Summarise A and B in parallel: 100 words each, same structure (who, claim, evidence, caveat). Output side by side. Highlight where their conclusions diverge.9. Voice-preserving summary
Summarise this in 300 words while preserving the author's voice. Here is a sample line in that voice: [paste one sentence the author actually wrote]. Don't flatten it to neutral encyclopedia tone.Swap: the sample voice line.
10. Headline plus one-line summary
Write 5 headlines (10 words or fewer each) and 5 one-line summaries (30 words or fewer each) for this piece. Vary the angle: outcome, problem, contrast, surprise, instruction. No clickbait.11. Truth-preservation check
Audit this summary against the source. Flag: (1) Claims in the summary that are not in the source, (2) Important caveats dropped, (3) Hedges removed where they mattered, (4) Numbers changed. Output a list of fidelity issues.Run this on the summary the model just gave you. It catches the hallucinated stat and the dropped “in some cases” that turn a summary into a misquote.
12. Audience-shifted summary
I have a summary written for engineers. Re-summarise the same facts for: (a) product managers, (b) executives, (c) end users. Each version 100 words. Keep the truth identical; shift only which facts get emphasised.Common mistakes
- Vague audience, so the output reads generic.
- No tone anchor, so every variant comes back tasting the same.
- No constraints — without a word cap and a banned-phrase list, the model pads.
- Skipping examples, even though they are the strongest signal for voice.
- Shipping the first draft. The model lands on the safe middle; the value is in your edit.
- Leaving in AI tells (“In today’s fast-paced world”).
- No fact-check pass. Run template 11 every time.
How to push results further
- Give 1-2 real tone examples instead of adjectives.
- Constrain ruthlessly: word count, format, banned phrases.
- Read the output aloud before publishing — clunky lines reveal themselves instantly.
- Cut adverbs and adjectives that do not carry weight.
- Use the model for drafts 1 and 2; your edit is draft 3, and draft 3 is what ships.
- Anchor the prompt in one real person from your audience, by name if you have one.
- Test the headline standalone — it has to work without the summary under it.
FAQ
- How long should the summary be?: Match the channel. A push notification gets one line; an executive brief gets three; a newsletter intro gets a paragraph. Set the cap explicitly or the model defaults to long.
- Can AI write the whole thing?: Use it for the first two passes, then edit. Across all 12 templates, the human edit is where fidelity and voice come from.
- Which model handles a long PDF or transcript best?: A 1M-token model — Claude Sonnet 4.6, Gemini 3.1 Pro, or Opus 4.7 for conflicting sources. Inside ChatGPT, Plus holds roughly 320 pages; the full 1M window needs the $200 Pro tier (as of June 2026).
- Should I publish without an edit pass?: No. The models are fluent and confident, but they drop caveats and occasionally invent numbers. Template 11 is your safety net.
- One voice or one per channel?: Keep a single brand voice and shift only the tone axes — formality, density, energy — per channel.
- Can I reuse these for other content types?: Yes. Swap audience, goal, and voice; the six-part structure is the reusable part.