AgentStack
SKILL verified MIT Self-run

Distill Knowledge

skill-siarhei-belavus-agent-public-distill-knowledge · by siarhei-belavus

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

No reviews yet
0 installs
12 views
0.0% view→install

Install

$ agentstack add skill-siarhei-belavus-agent-public-distill-knowledge

✓ scanned · ✓ verified — works with Claude Code, Cursor, and more.

Security review

✓ Passed

No issues found. Passed automated security review. · v0.1.0 How review works →

  • Prompt-injection patterns
  • Secret / credential exfiltration
  • Dangerous shell & filesystem operations
  • Untrusted network calls
  • Known-malicious package signatures

What it can access

  • Network access No
  • Filesystem access No
  • Shell / process execution No
  • Environment & secrets No
  • Dynamic code execution No

From automated source analysis of v0.1.0. “Used” means the capability is present in the source — more access means more to trust, not that it’s unsafe.

Are you the author of Distill Knowledge? Claim this listing to set pricing, connect Stripe payouts, and keep 70% of every sale.
Sign up to claim

About

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.
  1. 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.
  1. 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.
  1. 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.
  1. Handle drift explicitly.
  • If tracked AGENTS.md conflicts with code/tests, trust code/tests.
  • Propose update, supersede, or removal instead of preserving stale guidance.
  1. 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.
  1. Prune aggressively.
  • Append-only is forbidden.
  • Remove or supersede stale guidance when code/tests now encode it better.
  1. 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.
  1. 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

Do not

  • summarize the implementation
  • restate behavior already obvious from code/tests
  • promote every candidate just because it exists
  • use tracked AGENTS.md as a worklog or shadow spec
  • write AGENTS.override.md

Initiative mode

When initiative mode is active:

  • treat sync-initiative as the working-memory sync phase, not the durable distill phase
  • read the parent initiative packet plus relevant child ARTIFACT_CANDIDATES.md
  • read INDEX.md field Next durable distill: now | later
  • default to no_op mid-initiative unless the field says now and the residue is already stable/reusable
  • a child packet may live in an owner repo different from the parent initiative repo; durable distill should run in the repo that owns the residue
  • cross-repo initiatives may require separate fresh-context distill passes in different owner repos
  • at initiative close, run a final pass even if the correct durable result is still no_op
  • after any non-final pass or explicit no-op, reset Next durable distill: later
  • only the closeout pass that updates the parent initiative packet should write initiative Status: completed

Communication

Honor active caveman mode for user-facing replies per ../../references/communication-mode.md. Keep durable artifacts normal unless the human asks otherwise. Drop caveman for safety/clarity when needed, then resume.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

Install and usage instructions live in the source repository linked above.

Reviews

No reviews yet — be the first.

Versions

  • v0.1.0 Imported from the upstream source.