How to Write Prompts for Summarization That Don't Lose Key Details

Summarize this 80-page report produces a pleasant executive summary that omits the one number your boss will ask about. Compression without a contract loses details by design — unless your prompt defines what must survive.

2026 pattern (FutureAGI, Google Cloud, LangChain guides): pick chain type by document size, require citations, eval for faithfulness and completeness. Gemini 3.1 Pro and Claude Opus 4.8 handle long context; map-reduce still matters above window limits.

This guide covers three chain types, the detail-preservation contract, copy-paste prompts, before/after examples, eval checklist, and FAQ.

Before vs after: pleasant vs faithful

Weak prompt:

Summarize this quarterly board report for my CEO.

[80-page PDF pasted]

Typical output: fluent 3 paragraphs, round numbers instead of exact figures, action item owner names dropped, risk flagged in Q2 missing entirely, no source references.

Strong prompt:

AUDIENCE: CEO, 5-minute read

MUST KEEP: all revenue figures, Q3 risk flags, named owners of action items

OUTPUT: BLUF (3 sentences) | Key metrics table | Decisions and owners | Open risks

RULES: Every number cites section heading. If not in document, write NOT MENTIONED. Max 400 words excluding table.

Repeat MUST KEEP at end of prompt.

Expected output: exact revenue numbers with section refs, table with metric | value | source, every MUST KEEP item present or marked NOT MENTIONED.

Three summarization chains

Stuff — entire doc in one prompt. Best coherence and lowest cost. Use when doc fits context with room for output.

Map-reduce — summarize chunks in parallel, merge summaries. Use when doc exceeds context or corpus is huge. Risk: facts at chunk boundaries get dropped.

Refine — sequential rolling summary, each chunk updates running summary. Slower, better for narratives where order matters.

Default 2026 rule: try stuff first. Escalate to map-reduce only when needed.

Chain selection table

| Document size | Chain | Model note |

| Under 100k tokens | Stuff | Gemini 3.1 Pro single pass |

| 100k–500k tokens | Stuff or hierarchical map-reduce | Claude Opus 4.8 for fidelity |

| Over 500k or corpus | Map-reduce | Flash for map, Opus for reduce |

| Chronological narrative | Refine | Meeting transcripts, incident timelines |

| FAQ / encyclopedic | Map-reduce | Order-independent chunks |

The detail-preservation prompt contract

Every summarization prompt needs:

1. Audience — who reads this and why
2. Must-keep list — metrics, names, dates, decisions that cannot drop
3. Output schema — sections, word limits, bullet vs prose
4. Citation rule — quote or page/section reference per claim
5. Uncertainty rule — label inference vs source fact
6. Do-not-infer — explicitly ban filling gaps

Missing must-keep list is why summaries feel fine but fail review.

Stuff prompt (single pass)

AUDIENCE: [e.g. CFO, 5-minute read]

DOCUMENT: [paste or attach]

MUST KEEP: all revenue figures, Q3 risk flags, named owners of action items

OUTPUT:

1. BLUF (3 sentences)
2. Key metrics table (metric | value | source section)
3. Decisions and owners (bullets)
4. Open risks (bullets)

RULES:

• Every number cites section heading or page
• If not in document, write NOT MENTIONED — do not infer
• Max 400 words excluding table

Repeat MUST KEEP at end of prompt for long docs (lost-in-the-middle mitigation).

Worked output description: BLUF cites Q3 revenue exactly as in doc, table has 8 rows each with section ref, risks include supply-chain item from Section 4.2.

Map prompt (per chunk)

You are summarizing chunk [N] of [total] from a larger document.

CHUNK:

[text]

Extract ONLY:

• Facts with exact quotes for numbers and dates
• Named entities and decisions
• Open questions raised

Format: bullets. Tag each bullet [chunk N].

Do not synthesize across chunks. Do not omit numbers.

Chunk settings: 2-4k tokens with 10-15% overlap between chunks to preserve boundary context.

Example bullet: Revenue Q3: $4.2M (+12% YoY) — quote: "Q3 revenue reached $4.2 million" [chunk 3]

Reduce prompt (merge partials)

PARTIAL SUMMARIES:

[paste all chunk summaries]

Merge into one summary for [audience].

Deduplicate repeated facts. Preserve ALL unique numbers — if two chunks conflict, report both with sources.

OUTPUT SCHEMA: [same as stuff prompt]

MUST KEEP: [repeat list]

Do not introduce facts not present in partial summaries.

Run reduce on Claude Opus 4.8 or GPT-5.5 — not Flash — for high-stakes merges.

Refine prompt (narrative docs)

CURRENT SUMMARY:

[running summary]

NEW SECTION:

[next chunk]

Update summary to include new section. Preserve prior facts. Add new facts with section reference.

If new section contradicts prior summary, note contradiction explicitly.

Use for meeting transcripts, sequential reports — not independent FAQ-style docs.

Refine costs more tokens but preserves causal chains map-reduce smears.

Citation formats

| Trust level | Prompt requirement |

| High (legal, finance) | Exact quote + section ID for every claim |

| Medium | Paraphrase + section reference |

| Low (internal skim) | Bullet facts, mark [verify] on uncertain |

High-trust: require verbatim quotes for binding numbers.

Add to any prompt: Numbers without citation = invalid output — revise.

Audience-specific output schemas

| Audience | Schema emphasis |

| Executive | BLUF + metrics table + decisions |

| Legal | Exact quotes, defined terms, party names |

| Engineering | Repro steps, error codes, version numbers |

| Customer success | Action items, owners, customer quotes |

Same document, different MUST KEEP lists — never one-size summary for all stakeholders.

Worked before/after: meeting transcript

Weak:

Summarize this 90-minute product meeting.

Output: generic themes (alignment, timeline concerns), no decision log, names attached to wrong action items.

Strong:

AUDIENCE: PM distributing notes to eng + design

MUST KEEP: every decision, owner name, date commitment, unresolved blockers

OUTPUT: Decisions (decision | owner | deadline) | Discussion themes (3 bullets max) | Open blockers

RULES: Quote exact deadline dates. NOT MENTIONED if owner not stated.

Output: 4-row decision table, 2 blockers with names, no invented consensus.

Worked before/after: legal contract summary

Weak:

Summarize this vendor agreement.

Output: paraphrases liability cap wrong by an order of magnitude, omits termination notice period.

Strong:

AUDIENCE: Finance VP pre-sign review

MUST KEEP: payment terms, liability cap exact amount, termination notice days, auto-renewal clause, governing law

TRUST: High — exact quote for every MUST KEEP item

OUTPUT: Clause summary table (topic | exact quote | section)

RULES: Do not paraphrase numbers or dates. NOT MENTIONED if clause absent.

Claude Opus 4.8 preferred for quote fidelity.

Hierarchical map-reduce (very large corpus)

When flat map-reduce reduce step bloated or lossy:

Layer 1: Map chunks → partial summaries

Layer 2: Reduce partials into section summaries (group by doc section)

Layer 3: Reduce section summaries → final summary

Use same MUST KEEP on layer 2 and 3. Gemini 3.1 Pro handles layer 1 at scale; Opus on final reduce.

Map-reduce vs refine — when to switch

Map-reduce: corpus search, feedback tickets, encyclopedic reports — order doesn't matter

Refine: story arc, chronological events, policy docs where later sections amend earlier

If reduce step repeats itself or drops middle sections → switch to refine or hierarchical map-reduce

Chunk overlap: why 10–15% matters

Facts spanning chunk boundaries get dropped without overlap.

Example: chunk 1 ends mid-sentence on Revenue Q3 $4.2M — chunk 2 starts with YoY comparison. Without overlap, reduce may keep only half.

Prompt fix: Include 200-token overlap; map prompt instructs tag facts that continue from prior chunk.

Eval checklist (before you ship summary)

• Faithfulness: every claim traceable to source?
• Completeness: every MUST KEEP item present?
• No new facts: anything invented?
• Conflicts: contradictions surfaced, not smoothed over?
• Citation gap: any number without source?

FutureAGI 2026 eval rubric: faithfulness, completeness, coherence, citation accuracy — run manually or with eval tools on high-stakes summaries.

Spot-check script: for each MUST KEEP item, search summary — if missing, fail and re-run with item repeated twice in prompt.

JSON output variant (pipelines)

For downstream automation with GPT-5.5:

OUTPUT: JSON matching schema: { bluf: string, metrics: [{name, value, source}], decisions: [{text, owner, source}], risks: [{text, source}] }

RULES: Same citation and NOT MENTIONED rules. No extra keys.

Use structured output mode — see Structured output prompting article.

Model routing

Gemini 3.1 Pro: multi-PDF stuff within 1M context; cost-effective map-reduce at scale

Claude Opus 4.8: highest fidelity single-pass synthesis; best quote accuracy

GPT-5.5: structured output summaries with json_schema for downstream pipelines

Gemini 3.5 Flash: chunk map step in parallel workflows — not final reduce on critical docs

Never use Flash for reduce on board, legal, or compliance summaries.

Common failures

Summarize without audience → wrong abstraction level

Map-reduce without overlap → lost boundary facts

Reduce step allowed to infer → hallucinated glue

No MUST KEEP → model optimizes for fluency over coverage

Treating summary as ground truth in RAG — keep source accessible

Single reduce of 50 chunk summaries → middle collapse; use hierarchical

Repeating MUST KEEP only at start on 100k+ docs → lost-in-the-middle drops items

Common mistakes

Asking for executive summary without defining executive — CFO vs COO need different tables

Rounding numbers for readability — banned in high-trust prompts

Merging map prompts that say summarize chunk — extraction prompts outperform generic summarize

Skipping eval because summary reads well — fluency correlates poorly with completeness

Same summary for internal and external audiences — redaction rules differ

PromptMake workflow

/text → paste MUST KEEP + doc metadata → generate map/reduce prompt set → run in your LLM → human spot-checks metrics.

Save MUST KEEP list as reusable snippet per report type (board, incident, contract).

FAQ

Stuff or map-reduce for a 40-page PDF?

Stuff if under context limit with room for output — usually yes on Gemini 3.1 Pro or Claude. Map-reduce if you also paste appendices pushing over limit.

How long should MUST KEEP list be?

5–15 items. Too long dilutes focus; too short misses critical fields. Group by category: financial, people, dates, risks.

What if document contradicts itself?

Prompt: Report both versions with section refs. Do not resolve contradiction.

Can I summarize PDFs with images/tables?

Vision-capable models (Gemini 3.1 Pro, GPT-5.5) — add: Extract numbers from tables and chart captions; cite page and figure number.

How to summarize Slack/email threads?

MUST KEEP: decisions, @mentions with commitments, links shared. Map by thread day or 50-message chunks with overlap.

Reduce hallucination in merge step?

Reduce prompt: Every sentence must trace to a [chunk N] tag in partials. Delete any sentence without trace.

Is summary good enough for RAG?

No — store summary for routing, keep chunks with embeddings for retrieval. Summary is not a substitute for source.

Multi-document stuff prompt

When summarizing 3–5 related docs (e.g. contract + amendments):

DOCUMENTS: [label each Doc A, Doc B with paste or attach]

MUST KEEP: [cross-doc items — e.g. final liability cap after amendments]

RULES: If later doc amends earlier, use latest value and cite both docs. Flag conflicts.

OUTPUT: Same schema as single-doc stuff prompt.

Use Gemini 3.1 Pro for multi-PDF stuff within context; map-reduce if combined size exceeds window.

Related articles

Lost in the middle — why repeat rules at start and end

Claude vs ChatGPT vs Gemini — long-context routing

Structured output prompting — JSON summary schemas

Positive framing — NOT MENTIONED instead of don't guess

Bottom line

Summarization prompts are contracts: audience, must-keep list, output schema, citations, no-inference rule.

Stuff when you can. Map-reduce when you must. Overlap chunks. Eval faithfulness and completeness before anyone acts on the summary.