Clarify

/clarify — Disambiguate Before Acting

This skill exists because passive prose rules ("use AskUserQuestion proactively") get skimmed past. /clarify is a structural forcing function: when invoked, you MUST stop, audit, and ask. No exceptions.

When this fires

Explicit triggers (user):

• /clarify (slash command)
• "any doubts?" / "are you sure?" / "what are you assuming?"
• "check with me first" / "before you start" / "wait — confirm"

Self-triggers (you): the moment you catch yourself writing or thinking any of these on a non-trivial task — STOP and run /clarify on yourself:

• "I'll assume X" / "probably wants Y" / "going with Z"
• "while I'm here I'll also fix..."
• "looks like the user means..."
• "I think the right approach is..."
• "this could mean A or B — I'll go with A"

Skip /clarify when: the request is a single unambiguous action ("create a task called X", "read this file", "what does Y do"). Don't ask for permission to do what was directly asked.

Procedure

Step 1 — Audit (silent, internal)

If self-triggered mid-stream: halt any in-flight tool calls before completing the audit. The point is to ask BEFORE the action, not narrate it after.

Scan the current request and your in-flight plan for items in these four buckets. Be exhaustive — better to surface 6 things and skip 4 than to surface 2 and miss the one that mattered.

Bucket A — Ambiguities in the request Things said that could mean two different things. Examples:

• "archive this workflow" → archive in the tool? in registries? both? delete?
• "fix the email" → which email? rewrite or just typos?
• "update the script" → update what behavior? replace it?
• pronouns with unclear antecedents ("them", "that one", "the new one")

Bucket B — Scope assumptions Auto-expansions you're inclined to do. Examples:

• "while I'm here I'll also rename X" / "I'll clean up the surrounding code"
• "this needs a new helper too" (does it?)
• "I'll add tests" (was that asked?)
• touching files outside the requested scope

Bucket C — Unilateral tradeoffs Decisions you're making for the user instead of with them. Examples:

• library choice (lodash vs native, axios vs fetch)
• file placement (utils/ vs lib/ vs colocated)
• naming (function name, env var name, table name)
• breaking change vs additive (do you preserve old API?)
• architectural pattern (singleton vs DI, sync vs async, polling vs webhook)

Bucket D — Irreversible-action risk Anything that can't be reverted via API:

• sending messages (email, Slack, chat)
• pushing/force-pushing, dropping tables, deleting branches/files
• creating public artifacts, publishing packages
• modifying shared infra (production workflows, CI/CD, DNS)
• spending money

Step 2 — Triage (silent)

For each item found, decide:

• Ask — genuine ambiguity, multiple valid answers, the user's preference matters → goes into AskUserQuestion
• State and proceed — you have a strong default, low downside if wrong → mention it in your text reply ("Going with X — say if you want Y instead") and proceed
• Skip — truly obvious from context, asking would be insulting

Bias toward asking. If you're 60/40 on whether it's obvious, ask. The cost of asking is one click; the cost of guessing wrong is rework + trust.

Calibration check. If everything triaged to State-and-proceed or Skip and the Ask bucket is empty, ask yourself why /clarify fired. If you self-triggered, you noticed something. An empty Ask bucket on a self-triggered run usually means you rationalized the ambiguity back into "obvious" — the exact failure mode this skill exists to prevent. Re-surface at least one item.

Step 3 — Ask via AskUserQuestion

Build ONE AskUserQuestion call with up to 4 questions. If you have more than 4 genuinely-ambiguous items, ask the top 4 and hold the rest until those are answered.

Question construction:

• Each question is one focused decision — don't compound ("which library AND should we test?")
• 2-4 options per question, mutually exclusive
• Mark your recommendation with "(Recommended)" and put it first
• Each option gets a description explaining the consequence/tradeoff
• For Bucket D items (irreversible): one option should always be "Don't do this — explain alternative"

Don't ask:

• "Should I proceed?" / "Is this OK?" — those are non-questions
• Things the user already specified — re-asking signals you weren't listening
• Things you can verify yourself with a Read or Bash call

Step 4 — Apply answers

Once the user answers, restate the resolved understanding in one short line ("Got it — X with Y, leaving Z alone"), then proceed. Do NOT re-ask the same question on the next turn.

If the user selected "Other" with free-text, treat it as authoritative even if it contradicts your recommendation.

Anti-patterns

• Asking after you've already done the thing — defeats the purpose. Audit BEFORE the irreversible action.
• Asking obvious things — "should the function be named addUser or createUser?" when there's a clear codebase convention (read the codebase first). Asking is not a substitute for thinking.
• One question with 12 options — split into focused questions of 2-4 options each.
• Burying the recommendation — if you have a clear default, lead with it and label it "(Recommended)". Don't make the user guess what you'd pick.
• Re-running /clarify in a loop — once the user answers, those answers stick for the rest of the task. Don't re-ask on the next turn.
• Asking instead of doing trivial verification — if you can answer it by reading a file, read the file. AskUserQuestion is for things only the user can decide.

Why This Exists

The rule "use AskUserQuestion proactively" is a common piece of guidance in agent prompts and personal instructions, but it gets skimmed past on multi-step tasks — agents default to "make reasonable assumptions and proceed" because that's faster and feels productive. The cost shows up later: rework, broken trust, work done in the wrong direction.

/clarify is a structural forcing function. When invoked (by the user or by an agent self-recognizing a trigger phrase), it commands a literal AskUserQuestion call — not a prose rephrasing. The skill exists because behavior change requires a tool-shaped intervention, not another paragraph of guidance.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: davidsimoes
• Source: davidsimoes/claude-skills-shared
• License: MIT

Install and usage instructions live in the source repository linked above.