Use when writing acceptance criteria for a task - express each as an observable Given/When/Then that QA can execute, including negative cases
日本語の概要は準備中です。原文の説明を表示しています。
Use when a screen design needs variants - how many, one document per variant, what makes two variants genuinely different, recommending one, how the human's choice is read, and narrowing to the chosen one in revision
インストールする前に、エージェントに与えられる指示の中身を確認できます。
Variants exist so the human can choose a direction with the alternatives in front of them — not to show effort. Two variants that differ only in colour make the human do the designer's job; five variants for a field change waste the review.
Core principle: Variants differ in structure, share the design system, are each complete in their own document, and arrive with a recommendation.
Rule variant-count decides: the number the human names (description, comment or task chat), exactly; otherwise two for a new screen or flow and one for a small change to an existing screen. A design system task has none.
Pick the one or two axes that matter for THIS screen's main task, and vary those:
| Axis | Variant A | Variant B |
|---|---|---|
| Hierarchy — what comes first | the summary (totals, what needs attention) | the list itself |
| Layout | single column | master–detail / split view |
| Disclosure | one long page | tabs, or a stepper |
| Density | comfortable rows with secondary lines | a compact table |
| Interaction model | edit inline | edit in a dialog, or on its own page |
Never a variant axis: palette, radius, shadow, icon set, font — those belong to the design system, and all variants use the same one. Copy is the same in every variant except where the structure forces a different label.
design: <screen> · <letter> — design: Invoices — list · A, design: Invoices — list · B (screen-mockup-html). The review page lays the documents side by side; one document holding every variant cannot be compared or chosen.<h1> and #notes: Variant A — table first, not Option 1.Say which variant you recommend and why — the brief's primary action, the audience's main task, consistency with the existing screens, and how much it reuses from the inventory — in the recommended document's #notes and in your summary comment, naming it by its document title. Then record the choice as ONE non-blocking product question with record_open_questions:
prompt: "Which variant of the invoice list should be built?"
kind: product, blocking: false
recommended_answer: "design: Invoices — list · A — table first: finance staff scan 50+ invoices a day, and it reuses DataTable as is."
The human compares the documents side by side on the review page — page by page, commenting on either — and chooses one per screen that has variants, each choice posting a comment whose first line is exactly Chosen variant: <document title>. A choice for one screen says nothing about another; a screen drawn once is never offered. Any lines after it are the human's note on the choice ("keep B's empty state"): act on it like a review comment — a request to take parts of another variant is a combination, so a new variant (see In revision). Read the choice in this order:
Chosen variant: <document title> naming one of that screen's documents — it overrides the question, answered or not.One variant, one document, design: <screen> · A, drawn next to what is there today: Today and Proposed frames side by side at the same width, so the reviewer sees the delta without hunting for it. Draw "today" from the code you read, labelled as such. With one variant there is nothing to choose: no variant question.
Once a variant is chosen, revise only its document, keeping its title, sections and ids. The other documents stay on the task as they were — a document cannot be deleted — and the hand-off names the chosen one, so they are never built. update or withdraw the variant question. A request to combine ("A's table with B's summary") produces one new document with the next letter (· C), drawn in full, named for what it is.
Chosen variant: comment names another document.まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Use when writing acceptance criteria for a task - express each as an observable Given/When/Then that QA can execute, including negative cases
日本語の概要は準備中です。原文の説明を表示しています。
Use when the diff adds or changes an endpoint, resolver, RPC, job or query that takes an object id, a role check, a request binding or a tenant filter - BOLA/IDOR, function-level authorization, mass assignment and tenant scoping
日本語の概要は準備中です。原文の説明を表示しています。
Use on every UI change - semantic HTML, labels for controls, keyboard-navigable dialogs/menus, visible focus, and never color as the only signal
日本語の概要は準備中です。原文の説明を表示しています。
Use when a task changes any screen, form, dialog, menu or control - Lighthouse/axe scan of the changed screens, a keyboard walk, and the thresholds that fail a task
日本語の概要は準備中です。原文の説明を表示しています。
How to work a task returned with review, QA or UAT findings. Use when a task is in need_revision or PR review comments are in your context.
日本語の概要は準備中です。原文の説明を表示しています。
Use when deciding whether a request needs an analiz task before implementation - the conditions that require the architect's analysis versus going straight to implementation
日本語の概要は準備中です。原文の説明を表示しています。