<< All versions

Skill v1.0.1

currentLLM-judged scan90/100
kronael/tools/distill

+1 new

──Details
PublishedOctober 21, 2026 at 05:14 AM
Content Hashsha256:ad5936d439ddc63a...
Git SHAd138bd6ad64e
Bump Typepatch
Compare with v1.0.0
──Files
Files (1 file, 3.0 KB)
SKILL.md3.0 KBactive
SKILL.md · 115 lines · 3.0 KB

version: "1.0.1" name: distill description: /distill — Use when compressing a long document, transcript, or codebase into its essence. NOT for skill creation (use learn) or recall (use recall-memories). when_to_use: "long article, codebase audit, multi-page spec, chat transcript, research dump; what is this in 5 words, give me the gist, boil this down" user-invocable: true


Distill — the 5/3 approach

Compression reveals truth: noise falls away, essence emerges. All state lives in .distill/ under the project root, so a run resumes from wherever it stopped.

Where it runs

A distillation reads a whole codebase or research set. Invoked in the main thread, ALWAYS run it in a background Agent(subagent_type: distill), the thin agent that loads this skill, with the subject and scope; present .distill/final.md when it returns. A subagent that loaded this skill runs the phases itself.

Start here — read the state, pick the phase

.distill/ holdsPhase
nothing1 — decompose
areas.md, no breadth/2 — breadth research
breadth/ populated, no L1.md3 — compress each area
L1.md, no L2.md4 — cross-compress
L2.md, no final.md5 — final output
final.mdpresent it

Phase 1 — decompose → areas.md

Read the codebase or topic. Identify 5 major areas of roughly equal weight and decide per area whether it needs web research, codebase reading or both — NEVER force one source on every area.

markdown
# Areas
1.[Name] - [description] - [source: code|web|both]

Phase 2 — breadth → breadth/01-name.md …

Per area, gather ~500 words of raw findings: read files and grep patterns for code; WebSearch/WebFetch for context, state of the art and comparisons for web; both for both. No compression yet. Run areas sequentially and save after each.

Phase 3 — compress each area → L1.md

Per area, three levels, each shorter than the last:

markdown
## [Area Name]
### Detailed (~200 words)
[What it does, why it exists, key decisions]
### Summary (~50 words)
[One paragraph]
### Essence (3-5 words)
**[Core of this area]**

Phase 4 — cross-compress → L2.md

From the five essences and summaries:

markdown
# Cross-Area Analysis
## Patterns (~100 words)
[Trade-offs, architectural decisions, recurring themes]
## Summary (~30 words)
[One paragraph for the whole]
## Essence (3-5 words)
**[What this IS]**

Phase 5 — final → final.md

markdown
# Distillation: [Subject]
## Essence
**[3-5 words]**
## Summary
[One paragraph from L2]
## Patterns
[Cross-cutting from L2]
## By Area
### [Area 1]: [3-5 word essence]
[Summary from L1]

Present it. Done.

Rules

  • ALWAYS read the existing state before starting and save to .distill/

after each phase.

  • NEVER compress without reading the source first.
  • NEVER repeat a phrase between compression levels; ALWAYS make each level

shorter than the previous.

  • ALWAYS make the essence quotable and true. No marketing language.
← v1.0.0All versions