kg-elicit

name	kg-elicit
description	This skill should be used when the user mentions tacit knowledge, repeated mistakes, gotchas, lessons learned, 비법, 주의사항, 실패 경험, or invokes /kg-elicit. Captures implicit knowledge as Experience first, optionally promotes to Heuristic only after explicit confirmation.
trigger	/kg-elicit

/kg-elicit — Extract Tacit Knowledge

Goal: turn what's in the user's head into a wiki page before it's lost.

Activate When

• User mentions tacit knowledge, repeated mistakes, gotchas, lessons learned
• Korean keywords: 비법, 주의사항, 실패, 반복된 실수, 명심
• During /kg-ingest, after the main flow, if trigger keywords appear in source or recent conversation
• During /kg-reflect, when an Experience is referenced but no page exists
• User explicitly invokes /kg-elicit [topic]

Do Not Activate When

• The event is concrete and recent (last session/today) → use /kg-postmortem (stricter template)
• User wants to query existing experiences → /kg-query
• User wants summary of tensions → /kg-reflect

Preconditions

• Wiki must be bootstrapped (/kg-init has run, wiki/ exists with .schema/ and content folders)
• If wiki doesn't exist, tell user to run /kg-init first and stop

When to run

• Explicitly: user types /kg-elicit [topic]
• Auto: during /kg-ingest, after the main flow, if trigger keywords appear in the source or recent conversation
• Auto: during /kg-reflect, when an Experience is referenced but no page exists

Dialog (Experience-first routing)

Default outcome is always an Experience. Heuristic is a second, separate step that only runs when the user explicitly confirms the lesson generalizes.

Pass 1 — File the Experience (always runs)

Ask sequentially (max 5 questions; skip any the user answers upfront):

1. What did you try that didn't work? — seeds body section Attempted
2. What actually happened? — seeds Outcome (keep verbatim error text if any)
3. Why did it fail (your current theory)? — seeds Root Cause
4. How did you get past it, or is it still open? — seeds Resolution
5. One-line lesson? — seeds Lesson

Write wiki/experiences/<YYYY-MM-DD>-<slug>.md from the experience template. Set epistemic_status:

• observed if the event is concrete
• inferred if reconstructed from memory
• hypothesis if the user is still unsure about root cause

Update wiki/experiences/_index.md, wiki/index.md, wiki/log.md, and wiki/hot.md.

Pass 2 — Optional Heuristic promotion (only after confirmation)

Ask: "Does this lesson apply beyond this one case? If yes, phrase it as a general rule. If unsure or 'maybe', we leave it as an Experience for now — you can always promote later."

Only if the user gives a phrased rule:

6. Under what conditions would a newcomer hit this? — Heuristic.applies_when
7. Are there cases where the rule does NOT apply? — Heuristic "Does Not Apply When"
8. Confidence: high / medium / low? — Heuristic.confidence

Then write wiki/heuristics/<slug>.md from the heuristic template:

• Set distilled_from: [[experience-slug]]
• Add relations: - {predicate: learned_from, target: "[[experience-slug]]"}

Update wiki/heuristics/_index.md, wiki/index.md, wiki/log.md, and wiki/hot.md.

graphify integration (when available)

If graphify-out/graph.json exists AND mtime < 7 days:

• After Heuristic creation: call get_neighbors(<related_concept>) to verify the new Heuristic links to existing concepts (auto-suggest 1-hop wikilinks)
• For Experience: optional shortest_path(<symptom>, <root_cause>) to verify the link makes structural sense
• If graph stale (≥7d) or absent, skip this enrichment (Heuristic creation is wiki-layer; graph isn't required).

After writing

• Do NOT rebuild the graph — elicited knowledge is wiki-layer only
• Append to wiki/log.md:

  ## [YYYY-MM-DD] elicit | <one-line thesis>
  - Experience: [[experience-slug]]
  - Heuristic: [[heuristic-slug]] (or "none — deferred")
  

Principles

• Don't fabricate. If the user is vague, ask one sharpening question. If still vague, file as epistemic_status: inferred with confidence: low.
• No leading questions. "What went wrong?" not "Was it X that went wrong?"
• One page per lesson. Don't pack multiple heuristics into one page.
• Experience-first. A single "specific or general?" gate over-files Heuristics. Routing through Experience first keeps the raw case intact and makes promotion an explicit, datable act.

Output Contract

Elicit result: PASS | PARTIAL

Captured:
- Experience: [[experiences/<slug>]]
- Heuristic: [[heuristics/<slug>]] | none (deferred or rejected)

Epistemic status: observed | inferred | hypothesis
Per-item confidence: high | medium | low (Heuristic only)

One-line lesson:
<lesson>

Updated:
- experiences/_index.md
- heuristics/_index.md (if Heuristic created)
- index.md
- log.md
- hot.md (Recent Activity)

Confidence: high | medium | low

Caveats:
- <vague answers | event still unresolved | no Heuristic | graphify-stale (enrichment skipped) | none>

Next command:
- <none | /kg-reflect (if 3rd+ elicit) | /kg-challenge "<lesson>" | /kg-update>

Exceptions and Escalation

• Wiki absent → stop and suggest /kg-init.
• User vague after one sharpening question → file as epistemic_status: inferred, confidence: low. Do not invent details.
• User says "maybe general" → do NOT create Heuristic. Keep as Experience only.
• Multiple lessons in one input → split into multiple Experience pages (one event = one page).
• Promotion to Heuristic requires explicit "yes" + phrased rule per Authority Matrix.

Quality Gates

Before final answer:

• Experience page created (always, even if vague)
• Heuristic only created with explicit confirmation + phrased rule
• Indexes updated
• log.md entry added