Surface patterns in accumulated knowledge. Propose what to keep, promote, compact, or dismiss. The human decides.

Distill is the complement to : salvage extracts learning from a single session; distill curates the corpus those extractions build over time. Without distill, metis accumulates but never compounds.

Invoke when:

Use distill (not

/salvage

) when the session went well. is for stopping because things went wrong. Distill is for pausing because things went right — or simply finished — and learnings are worth capturing before context is lost.

Do not use when: You're in the middle of execution. Distill is a pause point, not a mid-flight activity.

Distill leverages LLMs for what they're good at — pattern recognition, clustering, surfacing similar entries — while keeping judgment with the human.

LLMs do: find recurring themes, group similar entries, surface candidates. Humans do: decide what matters, what's worth promoting, what's stale or context-specific.

Auto-promotion is never correct. A theme proposal is not a guardrail until a human writes it.

Decide what corpus to work with:

Session scope: Review the conversation. What was learned? What assumptions were validated or invalidated? What constraints were discovered? What would be useful to know at the start of the next session?

Corpus scope (RNA): Call broadly. Cluster by semantic similarity. Identify:

Present each candidate group with four possible actions:

**Theme: [theme name]**
Entries: [list with source IDs and one-line summaries]
Suggested action: [Keep / Promote / Compact / Dismiss]
Reason: [why this action fits]
→ Your call:

Keep — leave as individual metis entries, no change. Promote — pattern is recurring and stable enough to warrant a guardrail. Distill drafts a stub; human approves the content (editing as needed), then an agent writes the file:

---
id: [slug]
outcome: [outcome-id]
severity: soft
title: [one-line constraint]
---
[drafted body — human refines, agent writes to .oh/guardrails/]

Compact — multiple entries say the same thing. Human approves a merged version; originals archived or deleted. Dismiss — stale, superseded, or so context-specific it misleads more than it helps.

Only write what the human approved. No auto-promotion, no auto-deletion.

Scope: [session | corpus — filtered by: outcome/phase/tag] Entries reviewed: [N]

[repeat for each proposal]

Reads the current conversation. Surfaces what was learned this session. Proposes metis entries for human review and approval. Works after any completed session — the positive complement to (which is for drift and failure).

Output: approved entries as markdown, ready to write to .

Corpus mode: across all accumulated metis (or filtered subset), cluster by semantic similarity, surface candidate groups for human review.

Output is PR-able: distill produces a set of proposed file writes (new guardrail stubs, compacted metis, removal notes) formatted as a markdown summary. The human creates a branch, writes the approved files, and opens a PR — or uses to create issues for each promotion. Curation becomes a collaborative act, not a solo one.

Filtering options: pass outcome ID, phase tag, or recency window to to narrow the corpus. Useful when the full corpus is large but only a domain slice needs curation.

Comes after: Multiple sessions of or that have accumulated metis. Or: end of any session where learnings are worth capturing before context is lost. Leads to: Cleaner results. Guardrail candidates for human authoring. A compacted session file that seeds the next session. Relationship to

/salvage

: Salvage is per-session extraction (especially from failure). Distill is corpus-level curation (independent of any single session outcome). They're complementary — salvage feeds the corpus; distill keeps it from becoming noise.

After distill, typically:

Remember: Distill is not synthesis. The LLM finds patterns; you decide what matters. A corpus that accumulates without curation becomes noise. A corpus that's periodically distilled becomes situated judgment — the kind that actually improves future sessions.