grill-me · plugin87 — скилл для Claude Code: бриф и установка
Скилл grill-me для Claude Code задаёт вопросы о проекте до начала разработки и записывает оставшиеся решения как допущения. Его запускают перед созданием интерфейса.
MIT ★ 3 216 обновлён 2026-11-10
/plugin marketplace add plugin87/ux-ui-agent-skills
/plugin install ux-ui-agent-skills@ux-ui-agent-skillsЧто делает grill-me
grill-me разбирает задачу до начала разработки интерфейса: проверяет проект и задаёт только вопросы, ответы на которые изменят результат. Скилл учитывает существующие брифы, тему, соседние экраны и компоненты, чтобы не спрашивать о том, что уже видно в репозитории.
Неотвеченные вопросы превращаются в письменные допущения с выбранным вариантом и ценой ошибки. Затем скилл создаёт BRIEF.md с целью, аудиторией, содержимым, ограничениями, платформой и критериями готовности. За один раунд задаётся не больше семи вопросов.
Для проверок нужен полный UX/UI Agent Skills kit. Если установлен только каталог скилла через npx [email protected] add, он работает в режиме рекомендаций: проверки не запускаются и вердикта нет. После брифа порядок такой: сборка, /gate, /critique; сам скилл качество не оценивает.
Какие задачи решает
- выяснить требования до начала сборки интерфейса
- зафиксировать допущения по нерешённым вопросам
- описать состояния контента и ограничения экранов в
BRIEF.md - согласовать критерии готовности до реализации
Примеры запросов агенту
- Разбери бриф для нового экрана и запиши результат в
BRIEF.md. - Перед сборкой интерфейса задай только вопросы, которые повлияют на результат.
- Проверь требования к теме, содержимому и критериям готовности до начала работы.
Как установить grill-me
1Через маркетплейс плагинов
/plugin marketplace add plugin87/ux-ui-agent-skills/plugin install ux-ui-agent-skills@ux-ui-agent-skillsКоманда ставит плагин ux-ui-agent-skills целиком, со всеми его скиллами
2Через Skills CLI (Vercel)
npx skills add plugin87/ux-ui-agent-skills --skill grill-me -a claude-code3Вручную: положить файлы
~/.claude/skills/grill-me/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/grill-me/
1Через Skills CLI (Vercel)
npx skills add plugin87/ux-ui-agent-skills --skill grill-me -a codex2Вручную: положить файлы
.agents/skills/grill-me/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add plugin87/ux-ui-agent-skills --skill grill-me -a cursor2Вручную: положить файлы
.cursor/skills/grill-me/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install plugin87/ux-ui-agent-skills .claude/skills/grill-me2Вручную: положить файлы
.github/skills/grill-me/SKILL.md1Через Skills CLI (Vercel)
npx skills add plugin87/ux-ui-agent-skills --skill grill-me -a gemini-cli2Командой платформы
gemini skills install https://github.com/plugin87/ux-ui-agent-skills --path .claude/skills/grill-meВручную: положить файлы
.devin/skills/grill-me/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о grill-me
Как установить grill-me в Claude Code?
Добавь маркетплейс командой /plugin marketplace add plugin87/ux-ui-agent-skills, затем установи плагин: /plugin install ux-ui-agent-skills@ux-ui-agent-skills.
Что делает grill-me?
Скилл изучает бриф и файлы проекта, задаёт важные для реализации вопросы и записывает оставшиеся решения как допущения в BRIEF.md.
Что нужно для проверок с grill-me?
Нужен полный UX/UI Agent Skills kit. Если скопировать только каталог скилла через npx [email protected] add, проверки не запускаются и скилл не выдаёт вердикт.
Исходный текст
SKILL.md · MIT
--- name: grill-me description: Interrogate the brief before a line is built. Ask only the questions whose answers change the work, and put every unasked decision on the record as a stated assumption. # A command the user runs. Converted from .claude/commands/ so it has a # CLAUDE_SKILL_DIR of its own and its paths resolve on every install route. # disable-model-invocation keeps the behaviour a command had: the user # starts it, the model never does, and its description stays out of the # listing budget the model's skill choice is drawn from. disable-model-invocation: true --- ## Requirements This skill needs the full UX/UI Agent Skills kit: the gates, the references and the design-system library. Install it with `npx [email protected] init` or as the Claude Code plugin. `npx [email protected] add` copies the skill folder and nothing else. Installed that way, this skill runs in **guidance-only mode**: no gate runs, so it gives no verdict. > **Step 0: is the kit here?** Check once: > `ls ${CLAUDE_SKILL_DIR}/../../../tokens >/dev/null 2>&1 && echo KIT_OK || echo KIT_MISSING` > On `KIT_MISSING`, open your reply with: "Guidance-only mode: the UX/UI Agent > Skills kit is not installed here, so no gate ran and nothing below is > measured." Then work from this file and general practice, under three rules. > Do not guess or cite the contents of a kit file you could not open. State no > measured number (no contrast ratio, no N/N, no "passes WCAG"): write "not > verified" where one would go. End with a line `Skipped (kit not installed):` > that names every gate script this file tells you to run, then the install > command above. The gates catch a screen built wrongly. Nothing in this repo catches a screen built correctly from the wrong brief, and that is where the expensive rework comes from: four eval runs scored 14/14 on the gates and still came back from `/critique` as rework, every finding a decision no gate can see. This command spends five minutes up front to remove that class of failure. Target: `$ARGUMENTS` (the thing about to be built; if that is not clear, that is itself the first question). ## 1. Read before you ask Asking for something the repo already answers is noise, and it teaches the user that answering you is a waste of time. Before writing a single question: - Look for an existing brief, `CLAUDE.md`, `BRIEF.md`, or a `reference/` folder. - Look for an existing theme: `${CLAUDE_SKILL_DIR}/../../../tokens/*.json`, `design-tokens.json`, `theme.css`. If one exists, the theme question is answered. Do not ask it. - Look at the neighbouring screens and components. Framework, conventions, and the component inventory are usually visible, not unknown. Every question you keep after this pass is one the code genuinely cannot answer. ## 2. The bank, grouped by what no gate can see Pick from these. Do not ask all of them. **Purpose and lead** - Who is this for, and what is the one task they came to finish? - What is the first thing their eye should land on? (Four equal cards means the answer is "nothing", and that is a composition bug before it is a design.) - What does success look like for them? What have they done when they leave? **Content reality** - What does this look like with zero items? With one? With four hundred? - What is the longest string a real user will put in here: a 60-character project name, a nine-digit amount, an email with no spaces to break? - Which numbers are real and which are placeholder? Placeholder that ships as real is a lie the user catches immediately. **Irreversible actions** - Which actions here destroy or cannot be undone, and what confirms them? (Destructive wears the danger variant in every place it appears: the trigger and the confirm dialog both. A blue Delete is an automatic fail.) - What happens when the action fails? What does the user read, and what can they do next? **Theme and platform** - New theme or existing? If new: which direction, and does dark mode ship too? - What is the narrowest width this must survive? (The kit's floor is 280px.) - Which framework is the deliverable, and does it need RTL, AA or AAA? **Scope and done** - How many screens or components exactly, and what is explicitly out of scope? - What has to be true before this is done: which gates, and is `/critique` part of the bar or not? ## 3. The rule on asking - **Maximum seven questions, asked in one round.** A drip of one question at a time is worse than a wrong assumption. - **Every question must change the work.** If both answers produce the same build, it is not a question, it is a default. Take it and say so. - **Never ask what you can choose well.** Taste is the job. Ask about intent, content, and constraint; decide the rest. ## 4. Everything unasked becomes a written assumption An assumption held silently is indistinguishable from a guess. Each one gets a line: what was assumed, the default taken, and what it costs if it is wrong. ``` Assumption: no dark mode requested -> shipping both anyway (the theme is already dual, so the cost of including it is zero and the cost of retrofitting is not). ``` ## 5. Output: a brief that can be checked later Write `BRIEF.md` next to the work, in this shape: ``` # <what is being built> Goal one sentence, from the user's side User and task who, and the task they came to finish Deliverables exact list; N screens means N screens Lead the one thing the eye lands on first Content reality empty / one / many, longest string, real vs placeholder Irreversible which actions, what confirms them, what failure reads like Theme source of truth, dark mode, brand direction Platform floor narrowest width, framework, RTL, AA or AAA Out of scope what this is deliberately not Assumptions each with its default and its cost if wrong Done when the gates that must pass, and whether /critique is part of the bar ``` This file is the thing `/critique` and a later reviewer argue against. A brief that lives only in the conversation cannot be checked, and so it will not be. ## 6. Then build, then prove it `/grill-me` -> build -> `/gate` -> `/critique`. This command removes ambiguity. It does not produce quality, it does not score anything, and it is not a substitute for rendering the work and looking at it.
© Thientan Soparat · MIT · plugin87/ux-ui-agent-skills
Похожие скиллы
Все скиллы по теме «Дизайн» →Сначала — grill-me от других авторов.