| name | ask-questions |
| description | Gathers structured input from the user through targeted questions with discrete options. Use when invoked via /ask-questions, or when the user says phrases like "ask me questions", "clarify first", "gather requirements", "make sure you have context", or "what do you need to know". |
Ask Questions
Read the project's agent instructions (e.g. AGENTS.md) before applying this skill to ensure conventions are followed.
When to apply
- User explicitly invokes
/ask-questions
- User asks you to clarify, gather requirements, or check assumptions before proceeding
- You face unknowns that would significantly change your approach
Do not use this skill for:
- Single clarification questions mid-task (just ask inline)
- Things you can determine by reading project docs, code, or git history
- Implementation details with a single obvious answer given established patterns
- Whether to run tests, lint, or verify (always do it)
Procedure
- Identify the topic from the slash command argument (e.g.
/ask-questions about the deployment strategy) or surrounding context.
- Categorize your unknowns using the tiers below to prioritize what to ask.
- Present structured questions with discrete options. Use a structured-question tool (AskQuestion, AskUserQuestion) if available; otherwise present numbered text options the user can reply to by number.
- Ask all questions at once unless a later question depends on an earlier answer. Batch to minimize round trips.
- Proceed after answers: take the most appropriate next action (implement, plan, explore). Only pause if you genuinely cannot determine what to do next.
Question categories
Prioritize questions by how much they change your approach:
Tier 1 -- Blocking unknowns: Choices that change the implementation approach entirely. Ask these first.
- "Should this be a new handler or an extension of the existing one?"
- "REST API or GraphQL?"
Tier 2 -- Preference unknowns: Either option works, but the user has an opinion. Ask after Tier 1.
- "Return plain text or structured JSON?"
- "Single PR or split by concern?"
Tier 3 -- Confirmable assumptions: Things you believe are true based on code/docs, but want to verify. Ask last, or skip if confidence is high.
- "I'll follow the existing error handling pattern -- correct?"
- "This should go in the utils package based on the existing structure -- right?"
Question design
- Match option count to the actual decision space. Binary questions get 2 options -- do not inflate to 3-5 artificially.
- For genuinely open-ended option spaces, include an escape hatch ("Other" or "Agent decides").
- Allow multiple selections when choices can co-exist (e.g. "Which concerns apply?").
- One sentence per question. Front-load the decision being made.
After receiving answers
- If answers contradict established project conventions, conventions win. Explain why and ask the user to confirm they want to override.
- If the user says "you decide" or "just do it," proceed with your best judgment and state the assumptions you are making.
- Do not re-ask questions the user has already answered, even indirectly.
Related skills
- After gathering answers, chain into
/make-todos to decompose the work or /explore-codebase to investigate the relevant code.
