windags-skill-cdm-elicitation

license	BUSL-1.1
name	windags-skill-cdm-elicitation
description	Runs CDM (Critical Decision Method) interviews to elicit the perceptual layer missing from science-only SKILL.md files. Activate when: Looking Back Q3 fires with scope≠narrow, a pattern_crystallization event is detected, a skill underperforms (expectancy_violation), or a novel task has no catalog match. Produces structured SKILL.md v2 metadata: recognition-cues, expectancies, decision-cues, adaptive-workarounds. NOT for: general Q&A, skill selection, DAG execution.
allowed-tools	Read,Write,Edit
category	Agent & Orchestration
tags	["skill-creation","cdm","elicitation","knowledge-capture","cci"]
pairs-with	[{"skill":"skill-architect","reason":"CDM output feeds skill-architect to produce the final SKILL.md draft"},{"skill":"windags-looking-back","reason":"Looking Back Q3 fires this elicitation; Q4 feeds taxonomy updates"}]
io-contract	{"kind":"structured","outputSchema":"./schemas/output.json"}
metadata	{"recognition-cues":["Looking Back Q3 returns q3_generalizable=true with scope=medium or broad","Session log shows 3+ consecutive failures on same subtask then success","A skill was injected but confidence score < 0.4 after execution","Task completes with BM25 top-score below 0.35 (no catalog match)"],"expectancies":["Interview produces at least 3 decision-cue entries mapping branch → cue → indicates","Recognition-cues describe the task description, not the task category","Adaptive-workarounds capture at least one departure from the documented procedure","Output maps cleanly to SKILL.md v2 metadata fields with no ambiguity"],"decision-cues":[{"branch":"Select pattern_crystallization interview (fail→succeed)","cue":"Session shows N failures then success on same subtask within one session","indicates":"Expert knowledge just crystallized — hot cognition, interview NOW"},{"branch":"Select pattern_extension interview (novel success)","cue":"Task completed with high confidence AND BM25 top-score below 0.35","indicates":"New pattern not in catalog — extract recognition boundary via contrast"},{"branch":"Select expectancy_violation interview (skill underperformed)","cue":"Skill was injected AND confidence < 0.4 AND which_expectancy is known","indicates":"Skill's causal model is wrong or its NOT-FOR boundary needs extension"}],"adaptive-workarounds":["When the session expert (LLM) says 'I don't know why it worked' — use the timeline probe before the decision-point probe, the sequence reveals the cue","When interview produces abstract answers — ask 'What does the task description look like when you recognize this?' not 'When do you use this skill?'"],"execution-pattern":"sequential","needs-cdm":false}

WinDAGs Skill CDM Elicitation

Runs Critical Decision Method interviews to extract the perceptual layer that makes skills expert-level rather than just procedural (science-only).

When to Activate

Activate on:

• "Looking Back Q3 fired" / "Q3 scope=broad" / "Q3 scope=medium"
• "pattern crystallization" / "skill needs CDM" / "elicitation interview"
• "skill underperformed" / "expectancy violation" / "which expectancy failed"
• needs-cdm: true in any skill's metadata

NOT for:

• General skill writing (use skill-architect)
• Post-execution quality review (use windags-looking-back)
• Skill selection during DAG planning (use skill-selector)

Three Interview Protocols

Choose based on the trigger type. Each targets a different knowledge layer.

---

Protocol A: Pattern Crystallization (fail→succeed)

When: 3+ consecutive failures then success on the same subtask, within one session.

Why this protocol: Hot cognition. The expert (LLM session) just solved a hard problem. The solution path is maximally articulable RIGHT NOW. This is the moment Klein calls "crystallization" — a new entry is being added to the pattern library.

Probe sequence:

1. Timeline reconstruction (extracts the contrast cases automatically):

   "Walk me through the attempts that failed. What did you try first? Then what?"

2. Switch-point probe (extracts the key decision cue):

   "At attempt [N-1], what changed that made you try something different? What were you noticing or observing at that point?"

3. What-if probe (extracts the implicit prerequisite — the thing that would have let you succeed on attempt 1):

   "If you had known [X] from the start, would attempt 1 have worked? What is X?"

4. Cue inventory probe (extracts the recognition signal):

   "What does the task description look like — what are you reading or detecting — when you know this approach is the right one?"

5. Workaround probe (extracts the art layer — what wasn't in any existing skill):

   "Was there anything you did that you wouldn't find in a standard guide? Any departure from the usual approach?"