clarify

name	Clarify
description	Use when intent is ambiguous or conflicting—emit one <<<CLARIFY>>> block (option buttons in UI, no tools). Or user sends /clarify.
version	1.0.0
category	bundled
tags	["ux","ambiguity","clarification","dialog","choices","requirements"]
triggers	["which should I","pick one","either/or","not sure","ambiguous","conflicting requirements","what do you mean","unclear scope","CLARIFY","/clarify","choose between"]

Tool contract (read first)

Rule	Detail
No tools	Do not call any tool in the clarify turn — output plain assistant text only
Output	Exactly one <<<CLARIFY>>> … <<<END>>> JSON block on its own (runtime forwards it to the host UI)
After emit	Stop and wait for the user's choice as the next message
Manual test	Invoke via /clarify skill slash when installed, or trigger from ambiguous user messages

Hand off to workflow skills (open-web-research, task-execution, etc.) after the user picks — not before.

When to Use

• Requirements conflict, scope is fuzzy, or multiple valid interpretations exist.
• User asks "which should I…", "pick one", or gives either/or options without choosing.
• One sharp question resolves the fork faster than parallel tool guesses.
• UI can render options (browser host parses <<<CLARIFY>>> markers).

Relation to other skills

Use clarify first when the user’s intent is still unresolved (conflicting requirements, unclear fork, unknown deliverable). Once the question is scoped, follow the chosen workflow—e.g. open-web-research requires minimum search/fetch before pivots; do not use clarify to postpone evidence after intent is clear.

Output Contract

Emit exactly this block as plain assistant text (no tool call):

<<<CLARIFY>>>
{"question":"Which stack should we use?","options":["React + Vite","Plain HTML/CSS","Other"],"open_ended":false}
<<<END>>>

Rules:

• question: one concise sentence ending with ? when offering choices.
• options: 2–6 short strings when open_ended is false. Include Other only when genuinely needed.
• open_ended: true when typed prose is clearer than buttons — then options may be [].
• JSON must parse; use double quotes. Do not wrap in markdown fences inside the markers.

Then stop and wait — the host sends the user's choice as the next user message.

Pitfalls

• Asking overlapping questions in one turn instead of one clear fork.
• Using clarify to delay research after intent is already scoped.

Anti-patterns

• Do not ask three overlapping questions — one block only.
• Do not use this when you can reasonably default (note the assumption briefly instead).