Faithful paragraph summary
Summarize the source in 3–5 sentences for [AUDIENCE]. Preserve the central claim, supporting reason, and level of certainty. Do not add outside knowledge. If a point is unclear, label it uncertain. SOURCE: [PASTE TEXT]
ChatGPT • Claude • open models • prompts
AI Summary Prompts
Reusable prompt patterns for paragraphs, long text, PDFs, books, articles, email, transcripts, and decision records.
Prompt contract
Specify audience, fidelity, format, and uncertainty.
A strong summary prompt defines what the output is for, what must remain, what may be omitted, and how the model should handle missing information.
Layered long-text summary
Create three layers: (1) one-sentence overview, (2) five key points, and (3) a detailed brief with evidence references. Separate source facts from your inferences. End with omissions or uncertainties that a reviewer should check. SOURCE: [PASTE TEXT]
PDF with page references
Using the extracted PDF text, produce an executive summary with page references for every major claim. List unreadable or contradictory passages separately. Do not infer the content of missing tables, figures, or pages. EXTRACTED TEXT: [PASTE TEXT]
Book chapter map
For each chapter, record: purpose, main claim or event, supporting evidence, new terms or characters, and connection to the previous chapter. Then write a global summary that preserves the book’s argument or narrative arc. BOOK TEXT: [PASTE TEXT]
URL and article summary
Identify the article’s thesis, five supporting points, evidence used, named sources, limitations, and conclusion. Exclude navigation, ads, comments, and unrelated page elements. State when only a partial article is available. ARTICLE TEXT: [PASTE TEXT]
Podcast or video transcript
Create a timestamped brief with topic segments, speaker-attributed claims, key examples, decisions or recommendations, and unresolved questions. Mark transcription uncertainty rather than guessing. TRANSCRIPT: [PASTE TEXT]
Email thread summary
Summarize this email thread for a busy project lead. Include current status, decisions already made, requests, owners, deadlines, blockers, and unanswered questions. Distinguish the latest message from earlier context. THREAD: [PASTE TEXT]
Decision record
Turn the source into a decision record with: problem, options considered, criteria, decision, rationale, dissent or risks, owner, effective date, and review trigger. Do not invent missing fields; write “not stated.” SOURCE: [PASTE TEXT]
How to adapt the prompts
Replace bracketed fields and add source-specific requirements. For legal, medical, financial, or policy material, require direct verification of names, numbers, conditions, and quotations. For long sources, ask for a structure map before the final summary.
Model-neutral wording
These patterns work across ChatGPT, Claude, and many open models because they describe the task rather than relying on hidden product settings. Model behavior still varies, so test the prompt on representative examples and review the result against a written rubric.
Do not hide uncertainty
A summary system should be allowed to say “not stated,” “unclear,” or “transcript uncertain.” Removing those signals can make an output look cleaner while making it less trustworthy.
Related reading
Start with the text summarizer guide, then review the full article on prompting ChatGPT, Claude, and open models.