requirements-elicitlisted
Install: claude install-skill nolte/claude-shared
# Requirements Elicitation Skill
Operationalizes `spec/project/requirements-elicitation/` so a development request becomes a precise, confidence-scored requirement set instead of the agent's first plausible guess. The method is engineered against the default case: the user often **does not know exactly what they want**, **expresses it imperfectly**, or **is misread**. It drives understanding up to a measured threshold *before* anything downstream is built.
## Why this is a skill, not an agent
- **Mid-flow interactivity is the contract** — the spec mandates a one-question-at-a-time funnel interview with teach-back confirmation per requirement. Each turn shows one question, reads the user's answer, and lifts the lowest-confidence dimension; an agent's isolated, non-interactive context cannot run that dialogue.
- **Persistent on-disk artifact** — the result (requirement list + filled gap matrix + surviving assumptions) lives at `project/requirements/<slug>.md` and is read by downstream consumers (`roadmap-plan`, `feature-decompose`, `issue-orchestrate`). A skill owns persistent state; an agent returns a report and forgets.
- **`confirmed` vs `assumed` and teach-back need the user in the loop** — the spec forbids marking a requirement `confirmed` without an explicit teach-back confirmation; that signal can only be gathered interactively.
- Counter-dimension considered: a read-only agent could later score the gap matrix on a finished artifact (a `validate`-style audit), but the