Agent Skills: Ask Questions If Underspecified

Ask Questions If Underspecified

Step 1: Assess Underspecification

Identify which aspects are unclear or missing:

• Scope: What exactly should be included/excluded?
• Constraints: Performance targets, platform requirements, security rules?
• Acceptance criteria: How will the user judge success?
• Priority: Which conflicting options should win?

Do NOT ask about things you can determine by reading existing code and conventions.

Step 2: Triage Questions

Categorize each candidate question:

| Type | Ask or Decide? | | ------------------------------------------------- | --------------------------------------------------- | | Scope boundary (what to include) | Ask | | Non-functional requirement (perf/security target) | Ask | | Acceptance criteria | Ask | | Implementation approach | Decide yourself | | Code style / naming | Decide yourself (follow existing patterns) | | Nice-to-have polish | Decide yourself (proceed with sensible default) |

Step 3: Ask Concisely (Max 3 Questions)

For each blocking question:

1. State what you assume (best-guess default)
2. Ask the specific question in one sentence
3. Offer 2–3 concrete options when applicable

Format:

I'll [proceed with X default assumption]. Quick questions before I start:
1. [Specific question]? (default: [assumption])
2. [Specific question]? (default: [assumption])

</execution_process>

<best_practices>

1. Provide defaults with every question: Always state your best-guess assumption so the user can simply confirm rather than answer from scratch
2. Ask about outcomes, not implementation: "Should this support offline use?" not "Should I use IndexedDB or localStorage?"
3. Read the codebase first: Many apparent ambiguities are resolved by existing patterns — check before asking
4. Batch all questions in one message: Never ask one question, wait for answer, then ask another

</best_practices> </instructions>

User: "Add a search feature to the dashboard"

Good response:

I'll implement client-side search with debouncing and case-insensitive matching. Quick questions before I start:
1. Should search cover all columns or just the name/title column? (default: name/title only)
2. Should results filter in place or open in a new panel? (default: filter in place)

Bad response:

Before I start, I need to know:
1. What search algorithm should I use?
2. Should it be case sensitive?
3. What debounce delay?
4. Should I use a library?
5. How many results to show?

</usage_example> </examples>

Iron Laws

1. NEVER ask more than 3 questions at once — more than 3 questions causes decision paralysis; triage ruthlessly and decide the rest yourself.
2. ALWAYS provide a best-guess default alongside every question — a question without a default forces the user to think from scratch; a question with a default lets them confirm in one word.
3. NEVER ask about implementation details you can decide yourself — algorithm choice, code structure, naming conventions, and library selection are your decisions, not the user's.
4. NEVER block on nice-to-have requirements — identify what is truly blocking vs. what can be decided with a sensible default; proceed with the default and note your assumption.
5. ALWAYS read existing code before asking — check existing patterns, conventions, and related code first; most apparent ambiguities are resolved by looking at what already exists.

Anti-Patterns

| Anti-Pattern | Why It Fails | Correct Approach | | -------------------------------------- | -------------------------------------------------- | ---------------------------------------------------- | | Asking 5+ questions before starting | Paralyzes user; feels like interrogation | Triage to max 3 blocking questions; decide the rest | | Asking about implementation approach | That is the agent's job, not the user's | Ask about scope/constraints/acceptance criteria only | | Questions without defaults | User must think from scratch; slower feedback loop | Always state: "default: X — correct?" | | Sequential questioning (one at a time) | Creates a slow back-and-forth waterfall | Batch all questions into one message | | Asking things visible in the codebase | Shows insufficient research effort | Read existing conventions before asking |

Memory Protocol (MANDATORY)

Before starting:

cat .claude/context/memory/learnings.md

After completing:

• New pattern -> .claude/context/memory/learnings.md
• Issue found -> .claude/context/memory/issues.md
• Decision made -> .claude/context/memory/decisions.md

ASSUME INTERRUPTION: Your context may reset. If it's not in memory, it didn't happen.