Ask Clarifying Questions

Align your trajectory with the user's before code is written. The interview has two outputs: the load-bearing decisions, settled by the user, and a model of how the user decides, which you use for every decision the interview did not cover. Ask with the agent's interactive question tool when it is available; otherwise use concise prose.

Before Asking

• Use this skill only when it has already triggered and the request still has load-bearing ambiguity.
• Skip mechanical requests: typo fixes, obvious renames, or one-line bug fixes the user has already diagnosed.
• As a subagent without a direct user-answer channel, do not interview. Return the open questions and best-guess assumptions so the orchestrator can escalate.
• In non-interactive or headless runs, skip the interview, state the key assumptions, and proceed.
• Explore first, ask second. Read the brief, earlier turns, AGENTS.md or CLAUDE.md, open files, git status, and relevant diffs. Never ask what a read-only command or file read can answer; finding facts is your job, choosing between them is the user's.

Process

1. Map the decision space

Before the first question, privately enumerate the decisions the task involves and the plausible answers to each. Treat them as independent axes unless one answer removes another axis; most decisions are axes, not branches of a tree. Order the axes by how much settling each one shrinks the remaining space: an axis whose answer decides file layout, data model, or scope outranks one that decides a name. Predict an answer for every axis and note your confidence. Axes you can already predict with high confidence are not questions; they are the first entries in your model of the user.

The map stays private. Do not present it, and do not narrate the ordering.