pratiyush dev-spec-kit clarify — ≤5 questions, recommended-option, one at a time
Verification only works if the spec says something falsifiable. Before authoring criteria, find the
ambiguity and resolve it with the user — cheaply. Borrowed from spec-kit's /clarify: the model
pre-answers, the user mostly says "yes".
Detection passes (run all six over the intake/spec)
- Ambiguity — terms with more than one reading ("fast", "secure", "soon").
- Missing edge cases — unstated behavior on invalid input, empty/boundary, failure.
- Contradictions — two requirements that can't both hold.
- Under-specified acceptance — "works" with no observable outcome.
- Hidden assumptions — implied infra, auth, scale, data shape.
- Vague terms — undefined nouns the criteria will depend on.
Asking (RFC-2119)
- Ask ≤5 questions total, the highest-impact first, one at a time.
- Every question MUST carry a recommended answer (your best inference, marked "(recommended)") so the user can confirm with a word. Use the harness's question UI when available.
- You MUST NOT ask what the intake already answers, or what a sensible default settles — make the call, note the assumption, and move on (only escalate when guessing is costlier than asking).
- After each answer, fold it back into the intake/spec text immediately; don't batch.
Done
- Record the resolved decisions in the intake doc (and, if the project wants it auditable, as a
journaled note). Then hand to
dev-spec-kit-spec-author. A clarification you didn't write down didn't happen.