andurilcode/summarizer

Summarizer

Core principle: A summary is lossy compression. What you choose to lose defines whether the summary is useful or dangerous. The right summarization approach depends on what the reader will do with the summary — not on how long the source is. A summary for a decision-maker must lead with the conclusion. A summary for future reference must preserve retrievability. A summary for sharing must be self-contained. Different purposes demand different shapes.

The goal is not to make something shorter. The goal is to produce the smallest artifact that preserves the value the reader needs.

---

How to Execute This Skill

STEP 1 — Analyze the Source (internal only — NEVER include in output)

Before asking the user anything, classify the summarization task internally. This analysis is for your decision-making only — do not include it in your response to the user:

• Content type: article / document / transcript / conversation / code / research paper / book / multi-source collection
• Source length: short (<1000 words) / medium (1-5K) / long (5-20K) / very long (20K+) / multi-document
• Information density: sparse / mixed / dense
• Structure: well-structured / loosely structured / unstructured stream
• Contains: arguments / data / narrative / instructions / mixed

This drives the recommendation in Step 2.

---

STEP 2 — Recommend Style and Compression

Present the recommendation as a single interaction — one message with options the user can confirm or adjust.

Summarization Styles

| Style | How it works | Best for | Avoid when | |-------|-------------|----------|------------| | BLUF | Bottom Line Up Front. State the conclusion/recommendation/decision in the first sentence. Then provide supporting evidence in descending order of importance. Reader can stop at any point and has received the most important information first. | Decision-makers, busy stakeholders, status updates, any summary where "what should we do?" matters more than "what happened?" | Source is exploratory with no clear conclusion; reader needs to form their own opinion | | Key Points | Extract the N most important claims, findings, or ideas as a structured list. Each point is self-contained. No narrative flow — optimized for scanning. | Meeting notes, research papers, long reports, anything where the reader needs to quickly find specific information | Source is a narrative or argument where the logic between points matters as much as the points themselves | | Narrative | Rewrite the content as a coherent short-form narrative that preserves the logical flow and argumentation structure. Reads like a well-written paragraph, not a list. | Articles, essays, arguments, stories, anything where the reasoning chain matters — not just the conclusions | Source is a data dump or reference material where flow doesn't matter | | Briefing | Self-contained summary designed to be shared with people who haven't seen the source. Includes enough context that the reader doesn't need the original. Structured with: situation → key findings → implications → recommended actions. | Sharing summaries with teammates, leadership updates, cross-team communication, async briefings | Reader has access to the original and just needs a compressed version for themselves | | Progressive | Multi-layer summary. Layer 1: one sentence. Layer 2: one paragraph. Layer 3: full summary with key details. Reader chooses their depth. | Reference material, notes for future self, content you'll revisit later, knowledge management | One-time summaries where the reader won't return to the content | | Action-Oriented | Extract only what requires action: decisions made, action items, owners, deadlines, open questions. Everything else is discarded. | Meeting transcripts, planning sessions, retrospectives, any content where "what do we do next?" is the only question that matters | Content is informational with no actions (use Key Points instead) | | Comparative | Synthesize multiple sources into a single summary that highlights agreements, disagreements, and unique contributions from each. | Literature reviews, competitive analysis, multi-article research, any multi-source input | Single-source summarization |

Compression Levels

| Level | Ratio | Output | When to recommend | |-------|-------|--------|-------------------| | Headline | ~98% compression | 1 sentence | "What is this about?" — maximum compression, minimum nuance | | Snapshot | ~90% compression | 2-4 sentences | Quick orientation. Reader decides if they want to go deeper | | Standard | ~75% compression | 1-3 paragraphs | Default. Captures the substance without the detail | | Detailed | ~50% compression | Multiple paragraphs | When nuance, evidence, or caveats matter. Preserves reasoning chains | | Comprehensive | ~30% compression | Structured document | Long or complex sources where too much compression destroys value |

Fidelity Mode

Every summary makes a trade-off between readability and source accuracy. Make this explicit:

| Mode | What it means | When to use | |------|------------|-------------| | Faithful | Stays close to the source language. Preserves original terminology, attributions, and qualifications. Prioritizes accuracy over fluency. | Legal, medical, scientific, compliance — anywhere distortion is dangerous | | Rewritten | Rephrases freely in clear, simple language. May restructure arguments for clarity. Prioritizes readability over verbatim accuracy. | General content, articles, blog posts, meeting notes — anywhere clarity matters more than exact wording |

Default to Rewritten unless the content is in a high-stakes domain or the user asks for faithfulness.

Recommendation Logic

IF source is a meeting transcript or conversation:
  → Action-Oriented + Standard + Rewritten
  "Meetings are about decisions and next steps — I'll extract what matters."

IF source is a research paper or technical document:
  → Key Points + Detailed + Faithful
  "Dense technical content benefits from structured extraction with preserved precision."

IF source is an article, essay, or blog post:
  → Narrative + Standard + Rewritten
  "This reads as an argument — I'll preserve the reasoning chain."

IF user says "brief my team" or "share with leadership":
  → Briefing + Standard + Rewritten
  "I'll make this self-contained so they don't need the original."

IF user says "TL;DR" or "gist" or "quick summary":
  → BLUF + Snapshot + Rewritten
  (Explicit speed request — honor it directly)

IF user says "action items" or "what do we need to do":
  → Action-Oriented + Standard + Rewritten
  (Explicit action request — honor it directly)

IF multiple sources are provided:
  → Comparative + Standard + Rewritten
  "Multiple sources — I'll synthesize rather than summarize individually."

IF user wants to save for future reference or notes:
  → Progressive + Detailed + Rewritten
  "Multi-layer summary so you can scan at any depth later."

IF source is legal, medical, financial, or compliance-related:
  → Key Points + Detailed + Faithful
  "High-stakes content — I'll preserve source language and qualifications."

Incompatible Combinations

• Action-Oriented + Headline: Actions need enough detail to be actionable. Suggest Snapshot minimum.
• Comparative + Headline: Can't synthesize multiple sources in one sentence. Suggest Standard minimum.
• Progressive + Headline/Snapshot: Progressive IS multi-level — it already contains a headline. Use Progressive at Standard or Detailed.
• Faithful + Headline: Impossible to be faithful at 98% compression. Suggest Snapshot + Faithful.
• Briefing + Headline: Briefings need context to be self-contained. Suggest Standard minimum.

---

STEP 3 — Present Options to the User

Present the recommendation concisely. The user wants the summary, not a lecture about summarization theory.

Format:

I'd recommend a [style] summary at [compression] depth, [fidelity mode].

[1-sentence reason why this approach fits this content.]

Want me to adjust?
  Style: [recommended] / [alt 1] / [alt 2]
  Depth: Headline / Snapshot / Standard / Detailed / Comprehensive
  Fidelity: Faithful / Rewritten

If the agent platform supports structured input (e.g., ask_user_input), use it for frictionless selection.

Speed override: If the user pastes content and says "TL;DR" or "summarize this quickly", skip the interaction entirely and deliver immediately. Write 2-4 plain sentences or a short bullet list — no section labels, no template structure, no BLUF formatting. Match the casual tone of the request. Don't add friction to an explicitly fast request.

---

STEP 4 — Produce the Summary

Once style, compression, and fidelity are confirmed (or defaulted), produce the summary following the selected style's structure.

BLUF Execution

BOTTOM LINE: [Conclusion / recommendation / key finding in 1-2 sentences]

SUPPORTING DETAIL:
[Most important evidence or context — 1-2 sentences]
[Second most important — 1-2 sentences]
[Additional context if compression level allows]

SO WHAT: [Why this matters to the reader — 1 sentence]

Quality check: Cover the "BOTTOM LINE" section. If someone reads only that, do they have what they need to act? If not, your bottom line isn't bottom-line enough.

Key Points Execution

KEY POINTS FROM: [source title/description]

1. [Point — stated as a complete, self-contained claim. Not a topic label.]
2. [Point]
3. [Point]
...N points (scale to compression level: Snapshot=3, Standard=5-7, Detailed=8-12)

NOTABLE OMISSIONS: [Anything important you had to cut — 1 sentence]

Quality check: Read each point in isolation. Does it make sense without the others? If a point says "The authors also discussed methodology" — that's a topic label, not a key point. Rewrite as "The study used a randomized control trial with 500 participants across 12 months."

Narrative Execution

Write a coherent short passage that preserves the source's reasoning structure:

1. Open with the core thesis or finding
2. Walk through the key supporting arguments in order
3. Note the most important caveats or counterarguments
4. Close with the implication or conclusion

Quality check: Does the summary have a logical flow a reader can follow without jumping? If you removed the source entirely, could someone understand the argument from your summary alone?

Briefing Execution

BRIEFING: [Title]
Source: [what this summarizes]
Date: [when the source was created/published]
Prepared for: [intended audience, if known]

SITUATION: [Context the reader needs to understand why this matters — 2-3 sentences]

KEY FINDINGS:
- [Finding 1]
- [Finding 2]
- [Finding 3]

IMPLICATIONS: [What this means for the reader's work/decisions — 1-2 sentences]

RECOMMENDED ACTIONS: [What should happen next, if applicable]

OPEN QUESTIONS: [Unresolved issues from the source]

Quality check: Send this to someone who hasn't seen the source. Can they understand it fully? If they need to ask "but what was the original about?" — the briefing isn't self-contained enough.

Progressive Execution

## Layer 1 — One Sentence
[The single most important takeaway]

## Layer 2 — One Paragraph
[Core message expanded with 2-3 supporting points]

## Layer 3 — Full Summary
[Detailed summary preserving key evidence, caveats, and nuance.
 Structured with subheadings if the source has natural sections.]

Quality check: Each layer must be a complete, valid summary on its own. Layer 1 shouldn't require Layer 2 to make sense. A reader should be able to stop at any layer and walk away informed.

Action-Oriented Execution

DECISIONS MADE:
- [Decision 1 — stated clearly, with who decided if known]
- [Decision 2]

ACTION ITEMS:
| Action | Owner | Deadline | Status |
|--------|-------|----------|--------|
| [Specific action] | [Name] | [Date] | Open |

OPEN QUESTIONS:
- [Unresolved question 1 — who needs to answer it?]
- [Unresolved question 2]

PARKING LOT: [Topics raised but deferred — 1-2 sentences if any]

Quality check: Could someone who missed the meeting read this and know exactly what they need to do? If any action item says "Follow up on X" without specifying what "follow up" means concretely, it's not actionable enough.

Comparative Execution

SYNTHESIS: [Sources compared — list titles/descriptions]

CONSENSUS: [What all/most sources agree on]
- [Shared finding 1]
- [Shared finding 2]

DIVERGENCE: [Where sources disagree]
- [Source A says X; Source B says Y — the tension is about Z]

UNIQUE CONTRIBUTIONS: [What only one source covers]
- [Source C uniquely argues/provides...]

GAP: [What none of the sources address but probably should]

Quality check: Does each source get fair representation? Could a reader identify the distinct contribution of each source? If you removed one source, would the synthesis change — and can the reader see how?

---

STEP 5 — Offer Follow-up

After delivering the summary, offer exactly one follow-up based on style:

• BLUF: "Want me to expand on any of the supporting detail?"
• Key Points: "Want me to elaborate on any specific point?"
• Narrative: "Want me to compress this further or expand a particular section?"
• Briefing: "Should I adjust this for a different audience?"
• Progressive: "Want me to add a Layer 4 — your own executive annotation?"
• Action-Oriented: "Want me to draft follow-up messages for any of these action items?"
• Comparative: "Want me to deep-dive into any of the divergences?"

Offer naturally. Don't force if the user signals completion.

---

Calibration Rules

1. Compression is not deletion. Good compression preserves information density while reducing volume. After producing any summary, check: "Did I lose anything the reader would want back?"

2. Flag what you cut. At Standard compression or above, include a "NOTABLE OMISSIONS" note or "This summary does not cover..." to help the reader decide if they need the original.

3. Multi-source requires synthesis, not concatenation. Never summarize multiple sources independently and stack the results. The value is in cross-referencing — agreements, tensions, and gaps.