Discovery

Hold the problem space open long enough to map it. Close the session with a disposition — not a solution.

This skill is a conversation conductor, not a document filler. Its primary instrument is questions. Its primary discipline is sequence: problem-space questions strictly before solution-space questions. Its output is the exploration brief (assets/exploration-brief-template.md), the only artifact in the document pipeline that is permitted to contain contradictions, dead ends, and unresolved tension.

Workflow

1. Read references/facilitation-rules.md before the first question. It contains the question-sequencing discipline and the anti-sycophancy mechanics. Follow it during the whole session.
2. Open the problem space. If the user opened with a solution ("I want to build X"), excavate the problem behind it first: ask what goes wrong today without X. Offer this excavation; do not force it — the user can state that discovery happened elsewhere. If they confirm prior discovery, capture its conclusions into the brief and move to step 4.
3. Diverge. Alternate between mapping pains (who feels each, what it costs, which pains are symptoms of which) and generating solution directions. Never present one option when three are cheap to generate. For each option, produce the strongest argument for it and the strongest argument against it — including the user's favorite.
4. Surface the load-bearing beliefs. Extract assumptions the options rest on. Sort them: cheap to verify now, expensive (becomes a PRD open question), or load-bearing enough that being wrong kills a branch.
5. Converge on one thing only: the disposition. Proceed to PRD (name the direction and the runner-up), spike first (name the assumption and a cheap experiment), park (record why, so it reopens with context), or dissolve (no real problem — a success, not a failure).
6. Write the brief. Read references/ste-writing-rules.md first; all brief prose obeys those rules. Fill assets/exploration-brief-template.md from the session. Write it to a Markdown file and present it. Keep the rejected branches; they are the point.

Steps 2–4 loop freely. Step 5 happens once, at the end, and only when the map is good enough to choose a next action — not a winner.

Boundaries (refusals)