sum-summarize

name	sum - Summarize
description	Summarize anything — text, conversations, situations, projects, decisions, or experiences — using the interpretation that best matches what the user actually needs.
output	{"format":"prose"}

Summarize

Input: $ARGUMENTS

---

Step 0: Detect Interpretation

Before summarizing, identify what the user actually needs. The interpretations below were derived from three independent sources — the purposes people summarize for, the input types people bring, and the cognitive operations summaries perform — and validated by convergence across all three.

Interpretation	User's Purpose	What to Produce
Compress	Make something shorter while preserving content	Multi-level compression (1 sentence → 1 paragraph → 1 page)
Distill	Find the core insight in something complex	Irreducible core + supporting structure + what to ignore
Recap	Capture what happened in a temporal sequence	Decisions, actions, insights, open questions, emotional temperature
Status	Know where a project or effort currently stands	Current state, trajectory, blockers, next steps
Synthesize	Find patterns across multiple sources	Convergent themes, divergences, and the combined picture
Explain	Communicate something to a specific audience	Audience-adapted summary using their vocabulary and concerns
Retrospect	Wrap up a completed effort with lessons learned	What worked, what didn't, what to carry forward, what to stop
Extract	Pull specific categories of content from a body	Structured, ranked extraction of claims, actions, decisions, etc.

Detection Rules

• Long text, single source, "make it shorter" → Compress
• Complex argument, "what's the point?" → Distill
• Conversation, meeting, session log → Recap
• Ongoing project, "where are we?" → Status
• Multiple sources, "what's the picture?" → Synthesize
• "Explain this to [person/audience]" → Explain
• Completed effort, "what did we learn?" → Retrospect
• "Pull out the [X]" or "list all the [Y]" → Extract

Disambiguation

If ambiguous, ask using plain language (not internal terminology):

I can summarize this in different ways: (a) Make it shorter — keep the same content, just less of it (b) Find the main point — what's this really about? (c) List what happened — decisions, actions, open questions (d) Show where things stand — progress, blockers, next steps (e) Find patterns across sources — what themes recur? (f) Explain it for someone specific — adapted to their context (g) Wrap it up — what worked, what didn't, lessons learned (h) Pull out specific things — claims, actions, data, etc. Which fits?

Boundary Clarification

• Compress vs Distill: Compress makes the SAME content shorter. Distill finds the ONE core idea and structures everything around it. Test: if you expanded the summary back out, would you get the original (Compress) or a thesis paper (Distill)?
• Recap vs Status: Recap is BACKWARD-looking (what happened). Status is FORWARD-looking (where are we going). Recap answers "what did we discuss?" Status answers "are we on track?"
• Recap vs Extract: Recap always follows a temporal sequence (conversation, meeting, session) with preset categories. Extract works on any source and the user specifies what to pull out.
• Synthesize vs Compress: Compress works on ONE source. Synthesize works on MULTIPLE sources to find what they share and where they diverge.
• Retrospect vs Recap: Recap captures what happened. Retrospect evaluates what happened — what worked, what failed, what to change.

INTERPRETATION: [which one]
REASON: [why this matches]

---

Interpretation 1: COMPRESS

Produce the same content at multiple compression levels. Optimize for: fidelity (doesn't distort meaning) and compression ratio (genuinely shorter).

Step 1: Extract Claims

Read the input and list every distinct claim, fact, or point made.

CLAIMS:
1. [claim]
2. [claim]
...

Step 2: Rank by Importance

Claim	Centrality (1-5)	Novelty (1-5)	Total
[claim]	[score]	[score]	[sum]

• Centrality: How essential to the main point? (5 = IS the point, 1 = tangential)
• Novelty: Would the audience be surprised? (5 = very, 1 = obvious)

Sort descending.

Step 3: Compress at Three Levels

ONE SENTENCE — Top-ranked claim only. Complete, standalone, no jargon.

ONE PARAGRAPH (3-5 sentences) — Top 3-5 claims. First sentence = the one sentence. Last sentence = "so what" implication.

ONE PAGE (~250 words) — All claims scored 5+. Organized by logical grouping (not order of appearance). Key evidence for top claims. End with implications or next steps.

Step 4: Fidelity Check

Read each compression level back. For each, ask:

FIDELITY CHECK:
- Level 1: Does the sentence say what the original said, or did compression change the meaning? [FAITHFUL / DISTORTED — how]
- Level 2: Would someone who read only this paragraph draw the same conclusions as someone who read the full original? [YES / NO — what differs]
- Level 3: What was dropped? Does any dropped claim matter enough that its absence is misleading? [list dropped claims that matter]

If any check reveals distortion, revise that level.

Step 5: Verify

VERIFICATION:
- Level 1: Captures the single most important point? [YES/NO]
- Level 2: Someone could make a decision based on this alone? [YES/NO]
- Level 3: Nothing critical (scored 5+) missing? [YES/NO]
- Fidelity: No meaning distorted at any level? [YES/NO]

---

Interpretation 2: DISTILL

Find the irreducible core of something complex. Not shorter — clearer. Optimize for: insight (finds what matters) and signal-to-noise (eliminates what doesn't).

Step 1: Find the Core

The core is the ONE idea that all other ideas in the input serve. To find it:

1. State the thesis of each section or major point in the input:

   SECTION THESES:
   1. [section/point] → thesis: [what it argues]
   2. [section/point] → thesis: [what it argues]
   ...
   

2. Ask: which thesis do ALL the others support or elaborate? That is the core.

   • If one thesis is clearly served by the rest → that's the core
   • If two theses seem co-equal → find what they SHARE. The shared thing is the core.
   • If no hierarchy exists → the input has multiple cores. State each separately.

CORE INSIGHT: [The one thing this is really about, stated plainly]
CONFIDENCE: [high — clear hierarchy / medium — judgment call / low — multiple cores]

Step 2: Map the Supporting Structure

THE CORE IS SUPPORTED BY:
1. [supporting point] — [why it matters to the core]
2. [supporting point] — [why it matters to the core]
3. [supporting point] — [why it matters to the core]

THE CORE DEPENDS ON:
- [assumption or prerequisite the core requires]
- [assumption or prerequisite the core requires]

Step 3: Identify What to Ignore

SAFELY IGNORE:
- [element that seems important but is tangential] — [why it's not load-bearing]
- [detail that supports a supporting point but not the core] — [why it can be dropped]
- [context that's interesting but not structural] — [why it's noise]

Step 4: Distilled Output

DISTILLED
=========
CORE: [one sentence]
STRUCTURE: [3-5 supporting points, each one sentence]
DEPENDS ON: [1-3 assumptions]
IGNORE: [what you can skip without losing the essence]

---

Interpretation 3: RECAP

Summarize a conversation, meeting, session, or interaction. Optimize for: completeness of outcomes (nothing decided gets lost) and actionability (reader knows what to do next).

Step 1: Scan for Categories

Pull from the input:

DECISIONS MADE:
- [decision] — [who decided, if applicable]

ACTIONS IDENTIFIED:
- [action] — [owner, if applicable] — [deadline, if applicable]

KEY INSIGHTS:
- [something that shifted understanding]

OPEN QUESTIONS:
- [unresolved question]

WHAT CHANGED:
- [state before] → [state after]

DISAGREEMENTS / TENSIONS:
- [point of disagreement] — [current status: resolved/unresolved]

EMOTIONAL TEMPERATURE:
- [was this tense, smooth, energized, draining, etc.? one sentence]
- [why it matters: e.g., "tense — alignment on X may be fragile"]

Leave any category empty if nothing fits — don't force it.

Step 2: Narrative Recap

Write a 3-5 sentence narrative that covers: what the conversation was about, what was decided or discovered, what the emotional tone was, and what's still open.

RECAP
=====
[narrative paragraph]

DECISIONS: [list]
ACTIONS: [list with owners]
OPEN: [list]

---

Interpretation 4: STATUS

Summarize where a project, situation, or effort currently stands. Optimize for: honest assessment (not spin) and forward direction (what to do, not just what is).

Step 1: Current State

STATUS
======
WHAT: [what is this project/situation]
WHERE: [current state in one sentence]
PROGRESS: [what's been accomplished since the last meaningful checkpoint]
MOMENTUM: [accelerating / steady / stalled / uncertain — with one sentence of evidence]

Step 2: Blockers and Next Steps

BLOCKED BY:
- [blocker] — [what it prevents]

NEXT STEPS:
1. [most important next action]
2. [second most important]
3. [third]

RISKS:
- [thing that could go wrong] — [likelihood: high/medium/low]

Step 3: Verdict

OVERALL: [one sentence — is this on track, off track, or unclear?]
HONEST ADDITION: [one sentence the stakeholder needs to hear but might not want to]

---

Interpretation 5: SYNTHESIZE

Find the combined picture across multiple sources. Optimize for: pattern detection (what recurs) and divergence detection (where sources disagree).

Step 1: Source Inventory

SOURCES:
1. [source 1] — [brief description, 1 sentence]
2. [source 2] — [brief description, 1 sentence]
...

Step 2: Extract Key Points from Each

For each source, pull the 3-5 most important claims or findings.

SOURCE 1 KEY POINTS:
- [point]
- [point]
...

SOURCE 2 KEY POINTS:
- [point]
...

Step 3: Find Convergences and Divergences

Before reporting convergence, check source independence:

SOURCE INDEPENDENCE CHECK:
  Did sources share inputs, context, or instructions? [yes/no/partial]
  If yes: which context was shared? [describe]
  INDEPENDENCE VERDICT: [independent / partially independent / not independent]

  If NOT INDEPENDENT: Convergence across these sources may be ECHO, not signal.
  Multiple sources reaching the same conclusion from the same (possibly wrong)
  premises is not evidence — it's amplification. Flag any convergent finding
  where the converging sources shared the same context gap or assumption.

ALL SOURCES AGREE ON:
- [point that appears in 3+ sources] — INDEPENDENCE: [genuine / echo — shared assumption]
- [point that appears in 3+ sources] — INDEPENDENCE: [genuine / echo — shared assumption]

SOME SOURCES AGREE ON:
- [point] — found in: [sources] — absent from: [sources]

SOURCES DISAGREE ON:
- [topic] — [source X says A] vs [source Y says B]
  → Resolution: [which is more supported, or genuinely unresolved]

UNIQUE TO ONE SOURCE (potentially important or potentially noise):
- [point] — from: [source] — likely: [important insight / outlier / error]

Step 4: The Combined Picture

SYNTHESIS
=========
ACROSS [N] SOURCES:

THE COMBINED PICTURE: [2-3 sentence narrative of what the sources collectively say]

STRONGEST SIGNALS (appear in most sources):
1. [signal]
2. [signal]

OPEN CONFLICTS:
- [where sources disagree and it matters]

WHAT'S MISSING (none of the sources address):
- [gap]

---

Interpretation 6: EXPLAIN

Summarize something for a specific audience who doesn't have your context. Optimize for: audience fit (uses their frame, not yours) and standalone clarity (they don't need the original).

Step 1: Identify Audience and Gap

AUDIENCE: [who needs to understand this]
THEY KNOW: [what they already understand]
THEY DON'T KNOW: [what's missing]
THEY CARE ABOUT: [what matters to them — not to you]
FRAME: [how they think about this domain — their vocabulary, mental models, concerns]

Step 2: Bridge the Gap

Write the summary from THEIR perspective, using THEIR vocabulary, addressing THEIR concerns. Translate frames where needed (e.g., technical → business impact, research → practical implications).

EXPLANATION
===========
[summary written for the identified audience]

Step 3: Check

- Does this use their vocabulary, not mine? [YES/NO]
- Does this answer what THEY would ask, not what I'd want to tell them? [YES/NO]
- Could they act on this without asking follow-up questions? [YES/NO]
- Have I translated frames, not just simplified words? [YES/NO]

---

Interpretation 7: RETROSPECT

Wrap up a completed effort with lessons learned. Optimize for: honest evaluation (not self-congratulation) and transferable lessons (applicable beyond this specific effort).

Step 1: What Was the Effort?

EFFORT: [what was done]
GOAL: [what it was supposed to achieve]
OUTCOME: [what actually happened]
TIMEFRAME: [when it started → when it ended]

Step 2: Evaluate

WHAT WORKED:
- [thing that went well] — WHY it worked: [root cause of success]

WHAT DIDN'T WORK:
- [thing that went poorly] — WHY it failed: [root cause of failure]

WHAT WAS SURPRISING:
- [unexpected outcome] — LESSON: [what this teaches]

WHAT WOULD YOU DO DIFFERENTLY:
- [change] — BECAUSE: [what you now know that you didn't then]

Step 3: Carry Forward

RETROSPECT
==========
EFFORT: [one sentence]
OUTCOME: [one sentence]

KEEP DOING: [practices that worked and should continue]
STOP DOING: [practices that failed or weren't worth the cost]
START DOING: [new practices suggested by what was learned]
KEY LESSON: [the single most important takeaway, stated as a transferable principle]

---

Interpretation 8: EXTRACT

Pull specific types of content out of a larger body. Optimize for: completeness (nothing important missed) and usability (extracted items are organized, not just listed).

Step 1: What to Extract

Determine what the user wants pulled out. Common extraction targets:

• Claims — assertions, arguments, positions → rank by confidence level
• Actions — things to do, tasks, next steps → assign owners and deadlines
• Decisions — choices made or needed → note who decides and by when
• Questions — unresolved questions → rank by urgency
• Data — numbers, measurements, evidence → organize by category
• People — who's involved and their roles → map relationships
• Risks — potential problems mentioned → rank by severity × likelihood
• Requirements — things that must be true or done → classify as must/should/could

EXTRACTING: [what type]
FROM: [brief description of source]

Step 2: Extract

EXTRACTED [TYPE]:
1. [item] — [source/context]
2. [item] — [source/context]
...

Step 3: Organize

Apply the ranking method appropriate to the extraction type:

Extraction Type	Ranking Method
Claims	Confidence: high / medium / low / contested
Actions	Priority: must-do / should-do / nice-to-do + owner + deadline
Decisions	Status: decided / pending / blocked + decision-maker
Questions	Urgency: blocks progress / important / can wait
Data	Category grouping + outlier flagging
Risks	Severity × likelihood matrix
Requirements	MoSCoW: must / should / could / won't

ORGANIZED [TYPE]:

[HIGH PRIORITY / HIGH CONFIDENCE / MUST-DO]:
1. [item]
2. [item]

[MEDIUM]:
1. [item]

[LOW / CAN WAIT]:
1. [item]

---

Quality Criteria by Interpretation

Different summary types optimize for different things. Use the criteria for the active interpretation:

Interpretation	Primary Criteria	Secondary Criteria	Acceptable Trade-off
Compress	Fidelity (no distortion)	Compression ratio	May lose nuance
Distill	Insight (finds the core)	Signal-to-noise	May oversimplify
Recap	Outcome completeness	Actionability	May miss subtext
Status	Honest assessment	Forward direction	May feel blunt
Synthesize	Pattern detection	Divergence detection	May over-aggregate
Explain	Audience fit	Standalone clarity	May lose precision
Retrospect	Honest evaluation	Transferable lessons	May feel uncomfortable
Extract	Extraction completeness	Usability of output	May decontextualize

---

Output Format (All Interpretations)

Every output must include:

SUMMARY TYPE: [interpretation name]
INPUT: [1-line description of what was summarized]

[interpretation-specific output]

---

Depth Scaling

Depth	Compress	Distill	Recap	Status	Synthesize	Explain	Retrospect	Extract
1x	1 sentence only	Core only	Decisions + actions	State + next step	Convergences only	One paragraph	What worked / didn't	List only
2x	All 3 levels + fidelity	Full distillation	All categories + narrative	Full status + risks	Convergences + divergences	Audience-adapted + check	Full evaluation + carry forward	Grouped + ranked
4x	+ claim table	+ assumption testing	+ tensions + emotional temp	+ trajectory analysis	+ unique points + gaps	+ multiple audiences	+ root cause analysis	+ cross-references

Default: 2x.

---

Failure Modes

Failure	Signal	Fix
Wrong interpretation	Output doesn't match what user needed	Re-read Step 0; if ambiguous, ASK using the plain-language question
Compression without fidelity	Summary distorts or changes the original meaning	Run the fidelity check — compare summary claims to original claims
Can't find the core	Distillation has 5 "core insights"	There's only one core. Find the thesis all others serve. If truly no hierarchy, state that explicitly.
Recap without actions	Meeting summary is narrative only	Force the categories: decisions, actions, open questions
Status without honesty	"Everything is fine" when it isn't	Add the "honest addition" — what the stakeholder needs to hear
Synthesis without divergences	"All sources agree" (suspiciously)	Check harder — sources rarely fully agree. Find where they differ.
Echo mistaken for convergence	Multiple sources agree but shared the same inputs/context/assumptions	Run the source independence check. If sources aren't independent, their agreement proves nothing — flag convergent claims that rest on shared assumptions.
Explaining to yourself	Summary uses your jargon, not audience's	Rewrite from their vocabulary and frame, not just simpler words
Retrospect without discomfort	"Everything worked great"	If nothing went wrong, you're not being honest. Find the failures.
Extracting without ranking	50 items pulled out, flat list	Apply the ranking method for that extraction type. Raw lists aren't summaries.

---

Anti-Flattening Checks

Summaries are the most vulnerable output to voice collapse and cached takes — the summary often produces the most generic version of the content. Run these checks on every summary:

• Voice preservation: Does the summary preserve what's distinctive about the original? If the original had a strong voice, unusual framing, or specific terminology, the summary must retain that texture — not flatten it into neutral explanatory prose.
• Cached-take detection: Did the summary replace the original's actual claims with more familiar/generic versions? Compare: does each claim in the summary match what the original actually said, or did summarization quietly swap in the "standard" version?
• Specificity retention: Are concrete details, numbers, names, and examples preserved? Generic summaries drop specifics first — but specifics are often what made the original worth reading.
• Insight preservation: If the original made a surprising or counterintuitive point, does the summary still convey the surprise? Or did it normalize the insight into something obvious?

Pre-Completion Checklist

• Interpretation correctly identified (or asked using plain language if ambiguous)
• Boundary clarification applied (Compress vs Distill, Recap vs Status, etc.)
• Output uses the format and quality criteria for that interpretation
• Summary is genuinely shorter/clearer than the input (not just reformatted)
• The single most important thing is stated first
• Nothing critical was dropped (or drops are acknowledged)
• Interpretation-specific quality criteria met (fidelity for Compress, core-finding for Distill, etc.)
• Someone who reads only the summary could act appropriately
• Summary preserves the original's distinctive voice and framing (not flattened to generic prose)
• No original claims were quietly replaced with more "standard" versions

---

Integration

Use with:

• /sim → When you want to simplify (make easier) rather than summarize (make shorter/clearer)
• /tri → When items need urgency-based sorting, not summarizing
• /pri → When the summary reveals items that need prioritization
• /teach → When the goal is learning, not just understanding
• /reframe → When the summary reveals the framing is wrong
• /cmp → When you need a detailed head-to-head comparison (more thorough than Synthesize's divergence detection)
• /extract → When you need skills extracted from a prompt (different skill — extracts reasoning skills, not content)
• /phil → To derive requirements for what a summary should contain before summarizing
• /enough → To check whether the summary is thorough enough for its purpose