Skip to content
Agentmart
Submit a tool

distill-knowledge

Curate durable residue from a completed task packet into canonical tracked `AGENTS.md` files or choose an explicit no-op.

Skill by siarhei-belavus

IndexedMIT★ 2
Write a review

Install

Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.

npx agentmarket install skill/siarhei-belavus-distill-knowledge

Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.

Get the weekly diff

The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.

We use your email only for this newsletter. Privacy

About

Curate durable residue from a completed task packet into canonical tracked `AGENTS.md` files or choose an explicit no-op.

Use cases: productivity

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.

Tools
No tools found
Secrets
None asked for
Install scripts
None found

Network hosts it talks to 1

  • codeload.github.com · found in install command

Files it touches

  • ~/.claude/skills/distill-knowledge

What's inside

Skill name
distill-knowledge
Description
Curate durable residue from a completed task packet into canonical tracked `AGENTS.md` files or choose an explicit no-op.
License
MIT
Instructions
143 lines in SKILL.md · view on GitHub

Files (1)

  • SKILL.md

SKILL.md (first lines)

# distill-knowledge

This is the only workflow phase that may make semantic updates to tracked `AGENTS.md`.

`review-agents-tree` may auto-apply only trivial non-semantic hygiene fixes.

It must never write `AGENTS.override.md`.

This skill must run from a fresh context window. If the current context contains implementation/review backstory beyond what a fresh distiller should rely on, stop and restart before curating durable context.

Run it after implementation and both code-review phases are clean enough to freeze the task outcome.

## Read first

- `../references/task-packet-contract.md`
- `../references/persistent-artifacts-contract.md`
- `../references/review-phases-contract.md`
- `../references/final-state-authoring-policy.md`
- `../references/initiative-workflow-contract.md` when distilling an initiative outcome
- `../../references/communication-mode.md`
- relevant tracked `AGENTS.md` chain
- current task packet
- final relevant code and tests for any candidate under consideration
- final diff

If `AGENTS.override.md` exists locally, do not rely on it for canonical repository semantics.

## Distill workflow

1. Read the packet and final code/tests for durable residue only.
   - Focus on `PLAN.md`, review files, `AMENDMENTS.md`, `ARTIFACT_CANDIDATES.md`, the final relevant code/tests, and the final diff.
   - Verify candidates against the current code/tests, not just the diff hunks.
   - Ignore implementation chatter that does not survive contact with the final code.
2. Triage candidates.
   - Promote only candidates that are:
     - not recoverable reliably from code/tests
     - durable beyond the current task
     - costly to forget
     - not better encoded in code, names, or tests
   - Prefer no-op over weak prose.
3. Choose the right landing zone.
   - Use the nearest stable tracked `AGENTS.md` owner boundary.
   - Prefer the deepest correct tracked shard for one clear owner, but stop at the smallest stable common owner when the residue spans a broader module or several sibling owners.
   - In cross-repo initiative work, durable landing follows residue ownership, not automatically the parent initiative repo.
   - Do not create a new tracked shard without explicit human approval.
4. Choose the right section.
   - Every tracked `AGENTS.md` needs `## Scope`.
   - Keep `## Scope` lean and non-tautological: owner / applies-to / boundary clarification only. Do not merely restate the current directory path or obvious folder ownership.
   - `## Commands & Validation` for durable run/build/test/lint commands, first validation, or `done when` checks.
   - `## Entrypoints` only for root or hub-level start-here navigation that materially reduces search time. Deep leaf shards should usually omit it.
   - Avoid raw file-path links by default; prefer stable owner/module names or task-class wayfinding.
   - `## Constraints` for hard do/don't rules, invariants, or externally imposed constraints.
   - `## Decisions & Gotchas` for non-obvious rationale, tempting rejected paths, revisit triggers, or traps.
5. Handle drift explicitly.
   - If tracked `AGENTS.md` conflicts with code/tests, trust code/tests.
   - Propose update, supersede, or removal instead of preserving stale guidance.
6. Apply the approval rule.
   - Trivial navigation/hygiene updates may be applied directly.
   - Semantic add/supersede/delete requires explicit human approval before writing.
   - Sharding, scope changes, or meaning-changing conflict resolution requires explicit human approval.
7. Prune aggressively.
   - Append-only is forbidden.
   - Remove or supersede stale guidance when code/tests now encode it better.
8. Update `INDEX.md` with a short distill outcome note.
   - If the packet has no further implementation or review work after distill, it may be marked `done`.
9. Commit before reporting back when distill changed the workspace.
   - If you updated tracked `AGENTS.md`, packet artifacts, or other repo files, make a focused git commit before reporting to the lead.
   - Use a Conventional Commit message matching the actual distill change.
   - Include the commit hash and subject in the handoff.
   - If the correct distill result is `no_op`, do not create a dummy commit.

## Selection heuristics

Promote decisions when the future agent needs to know:

- why a strange choice is intentional
- which rejected path should not be resurrected casually
- what hidden external constraint shaped the code
- what revisit trigger would make the decision worth reopening

Promote entrypoints when the future agent would otherwise waste time rediscovering:

- where a recurring class of change starts
- which owner/module defines the behavior
- which first validation command gives fast confidence

Promote commands/validation when the future agent would otherwise guess:

- how to run the area safely
- which command gives the fastest useful signal
- what `done` means for recurring work in this boundary

## Commit-before-report rule

For distill work that changed the workspace:

- commit before reporting back to the requester
- include the commit hash and subject in the handoff
- prefer a focused commit over bundling unrelated residue edits together
- if the result is `no_op`, explicitly say no commit was needed

## Requester-response rule

Always report the distill result back to the requesting side.

- If the lead/orchestrator requested the distill pass, reply back to the lead/orchestrator.
- If a human directly requested the distill pass in this pane, reply back in this pane.
- Local pane output without an explicit reply to the requester does not count as completion.

## Output shape

Default distill outcome should say one of:

- `no_op`
- `updated_tracked_root_agents`
- `updated_tracked_local_agents`
- `proposed_semantic_agents_update_pending_approval`
- `pruned_stale_agents_guidance`

From siarhei-belavus/agent-public, MIT.

More from siarhei-belavus/agent-public

Signals

Stars
2
Forks
0
Open issues
0
Last commit (this folder)
Apr 29, 2026 (6 months ago)

From GitHub (fetched Oct 24, 2026).

Details

License
MIT
First seen
Oct 24, 2026
Last checked
Oct 24, 2026

More from siarhei-belavus

Cite this page

distill-knowledge skill. Agent Mart. https://agentmart.dev/skills/siarhei-belavus-distill-knowledge (last updated 2026-10-24)