Prompt Resumir em Bullets

Bullet-point summary prompts: turning long text into scannable insights

A bullet-point summary prompt asks a Large Language Model to compress a document — meeting transcript, article, research paper, book chapter, email thread — into a short list of high-information lines. The technique exploits one of the strongest capabilities of ChatGPT, Claude, Gemini and Llama: abstractive summarisation, where the model generates new sentences that capture meaning, rather than copying key phrases verbatim (extractive). The challenge is to keep the summary faithful while making it readable in 20 seconds.

A reliable prompt template

Summarize this {document_type} in 5-7 bullet points.
Each bullet:
 - starts with an action verb
 - max 12 words
 - captures one key insight
 - includes a specific data point if available

Document: {text}

The verb rule forces concrete claims (“Reduced churn 18% by simplifying onboarding”) instead of fuzzy descriptions (“Discussed churn improvements”). The 12-word limit caps cognitive load. The data-point hint pulls measurable evidence out of the source so the summary becomes evidence, not opinion.

Extractive vs abstractive, and chain summarisation

Extractive summaries pick original sentences (good for legal accuracy, weak for readability). Abstractive summaries rephrase — this is where LLMs shine. For long documents that bust the context window, use MapReduce chain summarisation: split into chunks, summarise each, then summarise the summaries. GPT-4 Turbo handles 128k tokens, Claude 3.5 Sonnet 200k+, Gemini 1.5 Pro up to 1M — longer windows mean fewer chunks but higher cost per call.

Summary levels and journalistic frameworks

• TL;DR — 1-2 sentences, social-media style.
• Executive summary — one paragraph, board-meeting style.
• Bullet points — 5-10 lines, the sweet spot for reports.
• Detailed summary — page or more for academic or legal use.
• 5W1H — Who, What, When, Where, Why, How. Journalistic baseline.
• BLUF (Bottom Line Up Front) — military style, conclusion first.
• Inverted Pyramid — newsroom standard, important info on top.

Output formats and verification

Bullets are not the only shape. Ask for a Markdown table for comparative summaries, a mind map for hierarchical ideas, or JSON when the output feeds another program. Whatever the format, remember that LLMs can hallucinate facts, dates and quotes — verify against the source for legal, medical or financial use. A useful trick: add “cite the line number for each bullet” so you can audit quickly. Role-based prompts also tighten the result: “Summarise as a senior consultant briefing the CEO” produces sharper, more decision-oriented copy than “summarise this”.

Tools that automate this in your workflow

• Otter.ai, Fireflies, Rev — meeting transcripts with auto-summary.
• Mem.ai, Notion AI — summary on top of your knowledge base.
• Readwise Reader — bullet summaries for articles and books.
• Glean, Hebbia — enterprise document summarisation with citations.

FAQ

Is it OK to copy the original words? Only if you need legal verbatim. For everyday use, abstractive (rephrased) summaries read better and compress further.

What is the ideal number of bullets? Five to seven. Below five you lose detail, above ten you exceed working-memory capacity (Miller’s 7±2).

Can the model summarise a mix of Portuguese and English? Yes — Claude 3.5 and Gemini 1.5 handle code-switching well. Pin the output language explicitly (“reply in English only”) to avoid mid-sentence switching.

What are the main risks? Lost nuance, minority opinions buried by majority views in meeting transcripts, and hallucinated specifics — especially numbers and dates. Always sanity-check the high-stakes bullets against the source.