Clarify Skill

Goal

At every step of any task, surface the questions that most reduce ambiguity before acting, and propose the questions the user may not have thought to ask. This generalizes the requirements-elicitation discipline (product/req-elicitation.skill, product/spec-writer.skill) beyond PRDs to every action — coding, refactoring, research, ops, review. The agent should be question-first whenever uncertainty is material, and silent only when the path is genuinely unambiguous and within autonomy bounds.

When to ask (uncertainty-aware)

Estimate your own uncertainty for the task/step and let it drive how much you clarify:

Uncertainty signal	What it looks like	Action
Multiple valid interpretations	The request maps to >1 plausible design/output	Ask — present the interpretations as options
Missing decision inputs	A choice depends on a fact you don't have (target, constraint, env)	Ask the specific fact
Retrieval miss	Memory/research found nothing for a key sub-claim	Ask or research, don't guess
Verifier/self-critique rejects	A self-review pass flags a likely error	Re-attempt or ask, don't ship
Irreversible / expensive / SOTA	Cost or blast radius is high	Always ask (regardless of autonomy)

Map this to the autonomy posture (GABBE_AUTONOMY):

ask — clarify on any non-trivial ambiguity.
hybrid (default) — auto-proceed when interpretation is unambiguous and the action is reversible and within budget; otherwise clarify.
auto — proceed silently for cheap/reversible work, but still always pause for expensive, SOTA, or irreversible actions (per AGENTS.md §9 and Article X).