Summarization prompts fail in two predictable ways: the model either repeats every detail (no compression) or drops critical information (over-compression). The fix is to specify exactly what you want — length, format, focus, and what counts as a key point.
Start With Length and Format
The single highest-leverage instruction in any summarization prompt is a concrete length target. "Summarize this" produces unpredictable results. "Summarize this in exactly 3 bullet points, each under 20 words" produces consistent, usable output.
Length can be specified in multiple ways depending on your downstream use:
- Word count: "Summarize in under 100 words"
- Sentence count: "Write a 2-sentence summary"
- Bullet count: "Give me 5 bullet points, max 15 words each"
- Character count: (useful for UI display) "Summarize in under 280 characters"
Format matters as much as length. A bullet list is not a paragraph. An executive summary is not a TL;DR. Name the format explicitly:
Summarize the following meeting transcript into:
1. A 1-sentence TL;DR
2. 3-5 bullet points of key decisions made
3. A "next steps" list with owner and deadline if mentioned
Transcript:
[transcript here]
This structure forces the model to separate decisions from next steps — something a generic "summarize" prompt never does.
Specify the Focus, Not Just the Topic
A 10-page contract and a 10-page research paper both require summarization, but they require different focus. Tell the model what kind of information is important for your use case.
Wrong: "Summarize this contract."
Right: "Summarize this contract focusing on: (1) payment terms and amounts, (2) termination clauses, (3) intellectual property ownership. Ignore boilerplate legal definitions."
The "ignore" instruction is underused. Models tend to include everything that seems important, which means legal boilerplate, background context, and methodology sections crowd out the actual decisions. Telling the model what to exclude is as powerful as telling it what to include.
Chain-of-Density: Progressive Compression
Chain-of-density is a technique from research that improves summary quality through iterative compression. Instead of asking for one summary, you ask for a sequence of increasingly dense summaries:
Summarize the following article using chain-of-density:
Step 1: Write a 5-sentence summary that captures the main points but may be verbose.
Step 2: Rewrite the summary in 3 sentences, adding 1-2 specific details while removing less important information.
Step 3: Compress to 2 sentences with maximum information density. Every word should carry meaning.
Article:
[article text]
Show all three versions.
This works because it forces the model to make explicit tradeoffs about what to keep. The intermediate steps also let you inspect where information is being lost — useful if the final version drops something critical.
You can also use chain-of-density in reverse: start with the compressed version and ask the model to expand with specific details, which often produces better targeted expansion than asking for a long summary directly.
Source Citation in Summaries
For documents where accuracy is critical — legal, medical, technical — ask the model to cite the source location for each summarized point:
Summarize the following document. For each point in your summary, include the paragraph number or section heading where that information appears.
Format:
- [Summary point] (Source: Section 3.2)
- [Summary point] (Source: Paragraph 7)
Document:
[document text]
This serves two purposes: it makes the summary auditable (you can verify claims), and it forces the model to ground each point in the actual text rather than inferring or interpolating. Grounded summaries are significantly more accurate for factual content.
The limitation is that models sometimes cite approximate locations. For high-stakes documents, always verify citations against the source.
Meeting Summaries vs. Document Summaries vs. Email Summaries
These three contexts require different prompt structures.
Meeting transcripts are chronological conversations with multiple speakers, often with tangents and repetition. The key information is rarely evenly distributed. Effective meeting summary prompts emphasize:
You are summarizing a meeting transcript. Focus on:
1. Decisions made (not discussions that led to decisions)
2. Action items: who will do what by when
3. Open questions that need follow-up
4. Blockers raised
Do not summarize general discussion or social exchanges. Skip anything that was ultimately overridden by a later decision.
Transcript:
[transcript]
Document summaries (reports, papers, contracts) have formal structure. The model can lean on section headings. The key prompt technique is asking for a structured output that mirrors the document's own structure:
Summarize this report. For each major section, write 1-2 sentences capturing the key finding or recommendation. Preserve the section structure.
If a section contains data or statistics, include the most important number.
Email summaries are short by nature, but email threads can be long. The critical distinction is summarizing the thread state (what is the current status?) vs. summarizing every email:
Summarize this email thread. Focus on:
- The original request or question
- The current status or resolution
- Any outstanding items
Do not summarize each email individually. Summarize the thread as a whole.
Handling Hallucination in Summaries
Models occasionally add details that were not in the source text, especially when the source is ambiguous or the model is "filling in" expected information. Two techniques reduce this:
Explicit grounding instruction: "Only include information that appears in the provided text. Do not infer, extrapolate, or add context from your training data."
Negative constraint: "If a specific detail (date, number, name) is not mentioned in the text, do not include it in the summary."
These instructions do not eliminate hallucination entirely, but they significantly reduce it for factual documents.
Calibrating for Audience
The same document needs different summaries for different audiences. A technical report summarized for engineers looks different than the same report summarized for executives. Include the audience in the prompt:
Summarize the following technical report for a non-technical executive audience. Avoid jargon. Focus on business impact and recommendations, not methodology.
vs.
Summarize the following technical report for a senior engineer. Include key technical decisions, performance numbers, and any architectural tradeoffs discussed.
The same source, two completely different useful summaries. Without the audience specification, the model picks one and you get lucky or you don't.
Practical Template for Most Use Cases
Here is a general-purpose summarization prompt that works well across document types:
Summarize the following [document type] for [audience].
Requirements:
- Length: [word count / bullet count / sentence count]
- Format: [bullets / paragraph / structured sections]
- Focus on: [specific topics or types of information]
- Exclude: [boilerplate / background / tangents]
- If a specific fact is not in the text, do not include it
[Document text]
Filling in each field forces you to think about what you actually need before you run the prompt, which is itself the most useful step in the process.
Keep Reading
- The Complete Prompt Engineering Guide (2026) — full foundation before specializing
- Chain-of-Thought Prompting with Examples — the reasoning technique that underlies chain-of-density
- Structured Output Prompting Guide — force JSON and schema-valid responses
Pristren builds AI-powered software for teams. Zlyqor is our all-in-one workspace — chat, projects, time tracking, AI meeting summaries, and invoicing — in one tool. Try it free.
