ask-design-questions

name	ask-design-questions
description	Ask the user a short, structured set of questions when intent, audience, style, or scope is unclear. Use before committing to a design direction — the app renders the questions as a real form with chips, textareas, and file uploads.

Ask design questions

Stop and ask when you are not confident about what the user wants. Guessing burns turns; one well-structured form usually unblocks everything.

When to use

Invoke this skill before writing any design file if you are uncertain about any of:

• Purpose — what is this for, who is the audience
• Content — subject, copy, names, brand
• Style — visual direction, tone, color
• Constraints — format, platform, size, deadline

If you only need one clarification, still use this skill. Do not embed questions in prose narration — the user answers prose with prose, which is messy to parse. A form is always better.

How to use

Write a single JSON file to .adits/questions.json at the project root. The file must conform to schema.json (sibling of this SKILL.md).

Then end your turn. Do not generate any other files in the same turn — the user can't answer questions and look at a draft at the same time.

What the app does with it

The file is a one-shot transport, not durable state. The moment the app sees it at the end of your turn:

1. It reads the JSON, validates against schema.json, attaches the payload to your turn, and deletes the file.
2. That turn is rendered as a form in the chat (not prose) — chips, textareas, dropzone. While this is the latest turn and unanswered, the composer greys out and a "Continue" button replaces Send.
3. When the user submits, the answers arrive as the next user turn. The form turn stays in chat history — scroll up and you'll see what was asked and what was picked.

Consequences for you

• The file is gone after one turn. Do not expect it to linger — if you end another turn without writing a fresh one, there's no form.
• Do not re-write .adits/questions.json on follow-up turns hoping to re-surface the same form. Each write = one new form turn. If you need new clarification later, write a new file with new questions.
• Refresh / project reopen does not re-pop the form. The chat transcript is the state; nothing ambient to auto-detect.

Authoring rules

• Keep it short. 3–6 questions is the target. 8 is a hard cap. If you have more, you are not asking questions — you are stalling.
• One concern per question. Never combine ("what is it for AND who is it for?"). Split into two.
• Lead with the question that most constrains the design. Purpose usually comes before style.
• Write in the user's language. If the user has been writing Chinese, the labels, hints, and options should be Chinese.
• Escape hatches on taste questions. For style/color/tone questions, set both allowExplore: true and allowDecideForMe: true — these questions have no wrong answer and the user may not have an opinion. Omit them on questions where you genuinely need a real answer (purpose, audience, content).

Schema cheat sheet

Top-level object:

{
  "title": "short heading for the whole form",
  "questions": [ /* 1–8 questions */ ]
}

Every question has: id (unique slug, [a-z0-9_]+), label, optional hint, and type — one of:

choice — chip row, single-select

{
  "id": "purpose",
  "label": "What is this webpage for?",
  "type": "choice",
  "options": ["Landing page", "Portfolio", "Company homepage"],
  "allowOther": true,
  "allowExplore": true,
  "allowDecideForMe": true
}

• options — 2–12 strings. Mutually exclusive; user picks one.
• allowOther — adds an "Other" chip with an inline text box.
• allowExplore — adds "Explore a few options" (you'll generate variants).
• allowDecideForMe — adds "Decide for me" (you pick).

text — free-text area

{
  "id": "topic",
  "label": "What's the subject or content?",
  "hint": "Product name, company name, key message — be specific.",
  "type": "text",
  "allowImageAttachment": true
}

• allowImageAttachment — adds a paperclip for inline image upload.

file — upload dropzone

{
  "id": "brand_assets",
  "label": "Have brand assets, logo, or reference images?",
  "hint": "Upload one if you have any; otherwise I'll use placeholders.",
  "type": "file"
}

Single file. No extra config.

The shape of the answers

After the user submits, the next user turn arrives as JSON in this shape:

{
  "answers": {
    "purpose": { "value": "Landing page" },
    "topic": { "value": "Adits is a tool for...", "images": ["files/hero.png"] },
    "brand_assets": { "file": "files/logo.svg" }
  }
}

Meta-selections show up as:

• { "value": "__other__", "text": "..." } for allowOther
• { "value": "__explore__" } for allowExplore
• { "value": "__decide__" } for allowDecideForMe

Read those and proceed — do not re-ask.