К основному содержанию
скилл CX Codex

start · donchitos — скилл для Claude Code: что делает и как установить

Онбординг для игровых проектов в Claude Code Game Studios. Запускает проверку репозитория, определяет стадию разработки и подбирает первый шаг.

Установить в Claude Code · Через Skills CLI (Vercel) Все способы и платформы ↓
npx skills add Donchitos/Claude-Code-Game-Studios --skill start -a claude-code
Открыть на GitHub

Что делает start

Скилл встречает разработчика после клонирования шаблона студии. Сначала скрипт bash .claude/scripts/project-state.sh проверяет репозиторий: ищет движок, исходный код, концепт-документы и архитектурные решения. Затем агент спрашивает, с чего начинается разработка: нет идеи, есть туманная задумка, готов чёткий концепт или уже написан код.

По ответам скилл вычисляет уровень строгости (modes.rigor) и стадию проекта (project.stage), после чего записывает их в project.yaml через вспомогательные скрипты. Запись происходит строго после подтверждения пользователем. Настройки вроде режимов ревью скилл не трогает, чтобы не перебивать ручные правила.

Скилл только предлагает следующий шаг вроде /brainstorm open или /setup-engine, но никогда не выполняет его сам. Если скрипты из каталога .claude/scripts/ отсутствуют, проверка состояния завершится ошибкой.

Какие задачи решает

  • определить текущую стадию разработки игрового проекта
  • выбрать уровень строгости документации и тестирования
  • подобрать первый шаг для работы над новой идеей или прототипом
  • проверить структуру папок и наличие дизайн-документов

Примеры запросов агенту

  • /start
  • Помоги начать работу над новой игрой
  • Я клонировал проект, с чего мне начать?

Как установить start

1Через Skills CLI (Vercel)

npx skills add Donchitos/Claude-Code-Game-Studios --skill start -a claude-code

2Вручную: положить файлы

~/.claude/skills/start/SKILL.md

Папка скилла целиком; для одного проекта — .claude/skills/start/

По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов

Из того же набора72

bug-report Скилл оформляет отчёты об ошибках по описанию или ищет баги в коде, а также помогает проверить исправления. Пригодится QA-инженерам и разработчикам игр для единого формата баг-репортов. CX Codex +1 473установок asset-audit Проверяет ассеты игры на соответствие правилам именования и лимитам веса, находя забытые файлы и битые ссылки в коде. CX Codex +1 469установок scope-check Сравнивает текущую работу с исходным планом, считает изменения объёма и предлагает, что сократить или перенести. Запустите его перед завершением спринта или этапа. CX Codex +1 460установок team-ui Скилл запускает пайплайн разработки игрового интерфейса от UX-спецификации до реализации. Распределяет задачи между проектировщиком интерфейсов и программистом. CX Codex +1 465установок prototype Скилл собирает черновой прототип игровой идеи и выдаёт вердикт PROCEED, PIVOT, KILL или NOT ASSESSED. Нужен до GDD. CX Codex +1 461установок reverse-document Разбирает уже написанный код или прототип и собирает по нему дизайн-док или ADR. Пригодится, когда фича есть, а описания к ней нет. CX Codex +1 461установок setup-engine Фиксирует движок и его версию в CLAUDE.md, а через WebSearch подтягивает справочники по API, если версия новее обучающих данных модели. CX Codex +1 453установок perf-profile Скилл ищет узкие места, сверяет замеры с бюджетами и ранжирует рекомендации. Нужен для профилирования игровых систем и кода. CX Codex +1 456установок

Другие версии start

Похожие

game-design-theory pluginagentmarketplace Скилл для ИИ-агента, который разбирает механику игры через MDA, психологию игрока и баланс. Пригодится, когда нужно понять, почему игра не держит внимание. CX Codex +1 2 500установок game-designer OpusGameLabs Скилл для геймдизайна: разбирает браузерную игру и доводит её вид — фон, частицы, анимации, переходы между сценами. CX Codex +1 1 900установок design-game-encounters MengTo Скилл для проектирования и настройки боёв в экшн-играх на Three.js: арены, враги, волны, фазы боссов. CX Codex +1 1 300установок roblox-player-psychology TabooHarmony Разбирает первые 60 секунд, награды, стрики, цены и шансы в кейсах. Пригодится, когда игроки не возвращаются или механика выглядит манипулятивной. CX Codex +1 1 200установок godot-genre-action-rpg thedivergentai Скилл для ИИ-агента, который ведёт разработку Action RPG на Godot: боевая система, статы, лут и прокачка по проверенным схемам. CX Codex +1 354установок start-learning rohitg00 Он проводит вводный опрос и placement quiz для курса AI Engineering from Scratch, а затем создаёт LEARNING.md с планом обучения. CX Codex +1 4 600установок

Рядом в теме «Геймдев и 3D: Unity, Unreal, Godot, Blender»

Всё по теме «Геймдев и 3D: Unity, Unreal, Godot, Blender» →

Вопросы о start

Как установить start в Claude Code?

Выполните команду npx skills add Donchitos/Claude-Code-Game-Studios --skill start -a claude-code.

Что делает скилл start?

Он проверяет проект скриптом project-state.sh, выясняет готовность идеи или кода и сохраняет начальные настройки в project.yaml. В конце он называет команду для следующего шага.

Запустит ли start генерацию концепта или настройку движка автоматически?

Нет. Скилл только определяет подходящую команду, например /brainstorm open или /setup-engine, и передаёт выбор разработчику.

Исходный текст

SKILL.md · MIT
---
name: start
description: "Guides a first-time user through onboarding and picks the right workflow. Use when the user is new, just cloned or installed the studio, has no idea or engine yet, asks where to begin, or wants to bring in an existing game."
argument-hint: "[no arguments]"
user-invocable: true
allowed-tools: Read, Glob, Grep, AskUserQuestion, Bash(bash .claude/scripts/project-state.sh*), Bash(bash .claude/scripts/set-stage.sh *), Bash(bash .claude/scripts/config-set.sh *)
model: sonnet
---

# Guided Onboarding

The entry point for a new user. It assumes no game idea, no engine and no experience: it reads
the project, asks where the user is starting from, shows what it assumes about the project, and
names the one next step. It never runs that step itself.

It writes these settings to `project.yaml`, all through helpers, and only after an explicit yes:
`project.stage` (with its legacy mirror `production/stage.txt`), `modes.rigor` and
`modes.automation`. It never writes `modes.review_mode` or `production/review-mode.txt`, and none
of the other knobs rigor fronts (`modes.workflow`, `docs.density`, `qa.level`,
`modes.story_granularity`, `team.size`): a written value pins that knob over the rigor expansion,
and a mirror file outranks it.

Ask with the question tool. Without one, list the options numbered and end your turn to wait. The
question is the last thing in the reply: no path, note or advice after it. A skipped or dismissed
question is not an answer: ask again in plain text.

The questions in Phases 2, 4 and 5 shape the project, so they are asked in every automation mode.
The engine is not one of them: `/setup-engine` chooses it. Phase 6's last question follows a
`modes.automation` already in `project.yaml`, per `.claude/docs/automation-modes.md`; with none
set, ask it.

## Phase 1: Read the project

Run `bash .claude/scripts/project-state.sh` and keep its rows. Do not show them yet; they check
the user's answer in Phase 2.

| Row | Use |
|---|---|
| `engine`, `code_root`, `source_files` | Is an engine set, and how much code is there. `source_files` of `unknown` means the code root could not be found: say that, never "no code" |
| `concept` | `design/gdd/game-concept.md` or `design/game-brief.md` exists |
| `gdd_files`, `adr` | Design documents and architecture decisions on disk (they set the stage in Phase 4) |
| `prototype`, `sprint`, `milestone`, `story` | Other existing work |
| `origin_url` | Contains `donchitos/claude-code-game-studios`, matched case-insensitively: the clone still points at the template repo, so a push or pull request would go there. A fork under another owner is the user's own |
| `stage` | The stage already recorded, if any |

Ignore `phase_inferred`: it guesses from file counts, and a project with code but no design
documents is not in Production. If the helper fails, say so and read the same facts by hand
(engine in `project.yaml`, the two concept files, the code root's files, `design/gdd/`).

The helper does not report rigor or automation. Read `project.yaml`, if it exists, and note
whether `modes.rigor` and `modes.automation` are set: a set value is kept, and its question is
not asked (Phases 4 and 5).

**Returning user** (an engine is set and `concept` is present): ask nothing. Say the engine, the
concept or brief found, and the review mode: `modes.review_mode` if set, otherwise the one
`modes.rigor` implies (`minimal` → `solo`, `standard` → `lean`, `full` → `full`). Suggest `/help` for the next step, add the
template-remote note if it applies, and stop.

## Phase 2: Ask where the user is

Ask with the question tool: "Welcome! Where are you starting from with your game?" Options:
- `A) No idea yet`: no game concept; explore and find what to make.
- `B) Vague idea`: a theme, feeling or genre ("something with space", "a cozy farming game").
- `C) Clear concept`: the core idea and genre are known, nothing is written down yet.
- `D) Existing work`: design documents, prototypes, code or plans already exist.

Wait for the answer. If it contradicts Phase 1, say what you found once and ask again: D on an
empty project ("this looks like a fresh template; would A or B fit better?"), or A with code on
disk ("I found code in `[code_root]`; did you mean D?"). If no option fits, let the user describe
their situation and route from that.

## Phase 3: Route

Name the immediate next step only. The full path waits for Phase 6, after the rigor is known.

- **A, no idea:** starting from zero is fine. `/brainstorm open` explores freely; `/brainstorm
  [hint]` starts from a word or theme. Next step: `/brainstorm open`, or a hint if one comes to
  mind. Describe it as "turns no idea into a written design the next step builds from"; how long
  that design is depends on the rigor, so leave that for Phase 6.
- **B, vague idea:** ask for the idea in a few words, take it as it is, and name
  `/brainstorm [their words]`.
- **C, clear concept:** ask for the concept in one sentence (genre and core mechanic), as plain
  text. Then ask: "How would you like to proceed?" Options: `Formalize it first` (`/brainstorm
  [concept]` writes it down) / `Jump straight in` (`/setup-engine` now, write the design after).
  The next step is their pick.
- **D, existing work:** say what Phase 1 found ("N source files under `[code_root]`, N design
  documents, engine [X / not set]"). Next step: `/setup-engine` when no engine is set and there is
  no more than a concept; otherwise `/project-stage-detect`.

Then say: "I'll lay out the full path in a moment."

## Phase 4: Show what you assume, then save it

Work out two values.

**Stage.** Paths A, B and C: `Concept`. Path D, from the Phase 1 rows:

| On disk | Stage |
|---|---|
| no engine, or nothing beyond a concept (`gdd_files` 0) | `Concept` |
| `gdd_files` 1 or more, `adr` 0 | `Systems Design` |
| `adr` 1 or more | `Technical Setup` |

**Rigor.** If `modes.rigor` is already set in `project.yaml`, keep it and say so. Otherwise infer
it from what the user said in Phases 2 and 3, using the signals in
`.claude/docs/settings-guidance.md` §3:

| The user's words mention | Rigor |
|---|---|
| open world, RPG, sim, sandbox, factions, crafting, economy, MMO, multiplayer, many systems | `full` |
| release, commercial, Steam, one clear core loop | `standard` |
| jam, weekend, prototype, first game, small, just trying | `minimal` |
| nothing above, or no idea yet | `minimal` (the default) |

A path B hint is still a description: "some kind of open-world survival sim" is `full`.

Show the card:

```
What I'll set up
  Stage:  [stage]  ([why, from what was found or picked])
  Rigor:  [value]  ([why, in the user's words] | already set)
          minimal: short brief, coarse stories, tests optional; about 4 steps to code
          standard: about 8 design documents before code; standard QA evidence
          full: everything in standard plus validation builds; evidence everywhere
  /settings changes either of these any time.
```

Then ask with the question tool: "Save these to `project.yaml` (and the stage to
`production/stage.txt`)?" Options: `Save them` / `Don't save`. Ask no question about the rigor
itself: the card shows it, and `/settings` changes it.

- **Save them:** run `bash .claude/scripts/set-stage.sh "[stage]"`, then, only when rigor was not
  already set, `bash .claude/scripts/config-set.sh modes.rigor [value]`. Check the read-back rows
  match. Say: "Saved. Rigor `[value]` sets six other settings; `/settings` shows them and changes
  any of them."
- **Don't save**, or anything other than an explicit yes: write nothing, keep the values for this
  run, and say they were not saved.

If a helper exits non-zero, show its output, say which setting was not written, and go on with the
value for this run.

## Phase 5: Ask how much to confirm

If `modes.automation` is already set, say "Automation is `[value]`" and go on. Otherwise ask:
"Last one: how much should I confirm with you as we work?" Options:
- `Guided (Recommended)`: I decide small things and still stop for big ones (scope changes,
  deletions, schema changes).
- `Collaborative`: I ask before each significant step and show drafts before writing.
- `Autonomous`: I proceed and log decisions, stopping only for the always-ask categories. Fastest,
  most tokens.

Mapping: `Guided` → `guided`, `Collaborative` → `collaborative`, `Autonomous` → `autonomous`. The
pick is the consent to write it: run `bash .claude/scripts/config-set.sh modes.automation [value]`,
unless the Phase 4 save was declined (then say it was not saved).

## Phase 6: Show the path

Say: "Here's your path at rigor `[value]`. Every skill still runs at any rigor; rigor changes
what's required, not what's allowed."

The path starts with the step Phase 3 named. When that was `/brainstorm` (paths A and B, and C's
`Formalize it first`), list `/brainstorm` first and `/setup-engine` second: each runs without the
other.

**`minimal`: 4 steps to running code**
1. `/setup-engine`: choose and configure the engine
2. `/brainstorm`: write the one-page `design/game-brief.md`
3. `/create-stories`: turn the brief's MVP list into stories (no epics or sprint plan needed)
4. `/dev-story`: the first line of game code

**`standard`**
- Concept: `/setup-engine` → `/brainstorm` → `/prototype` → `/art-bible` → `/gate-check`
- Systems Design: `/map-systems` → `/design-system` (once per system) → `/review-all-gdds` →
  `/gate-check`
- Architecture: `/create-architecture` → `/architecture-decision` (once per decision) →
  `/create-control-manifest` → `/architecture-review` → `/test-setup` → `/ux-design`
  (accessibility and interaction patterns) → `/gate-check`
- Pre-Production: `/ux-design` (HUD and key screens) → `/create-epics` → `/create-stories` →
  `/sprint-plan` → `/gate-check`
- Production: `/dev-story`

Each `/gate-check` moves the project one stage, and only when it passes.

**`full`:** everything in `standard`, plus `/vertical-slice` and `/playtest-report` in
Pre-Production and `/design-review` after each GDD.

**Path D with existing work** (stage `Systems Design` or later, or code on disk) gets this instead.
Print only the steps that apply, and for the brief only the one source that applies. At `minimal`:
1. `/project-stage-detect`: what stage the project is at and what is missing
2. `/adopt`: whether the existing files are in the format the skills read
3. `/setup-engine`: only when no engine is set; leave it out when `engine` names one
4. The brief, only if `concept` is absent: `/reverse-document design [code_root]` when the code
   root holds 10 or more source files (it writes the brief from the game), and then do not mention
   `/brainstorm` at all; `/brainstorm` only when it holds fewer

At `standard` and `full`, steps 1 to 3, then `/design-system retrofit [path]`,
`/architecture-decision retrofit [path]`, `/architecture-review` and `/gate-check`.

Present `minimal` as the documented floor, not a lesser mode, and name its trade once: no GDDs, so
design problems show up in code.

If `origin_url` pointed at the template repo, add: "Your clone's `origin` still points at the
CCGS template repo, so a push or pull request would go there. When you're ready, run
`git remote rename origin template`, then `git remote add origin <your repo URL>`. `UPGRADING.md`
pulls updates from `template`; skip its own `git remote add template` step, since that remote now
exists." Suggest it; do not run it.

Ask: "Would you like to start with [first step]?" Options: `Yes, start with [first step]` /
`Something else first`. Either answer leads to Phase 7, never to running the step.

## Phase 7: Hand off

Respond with exactly one line: "Type `[skill]` to begin." Do not start that skill or call the
skill tool for it, even after a yes: the user types it. If they picked something else, name the skill for that instead, the same
way.

Verdict: **COMPLETE**, the user is oriented and handed off. If the user stopped before Phase 7,
say which settings were saved and which were not.

© Donchitos · MIT · Donchitos/Claude-Code-Game-Studios

Смотрите также

Спонсоры проекта

Компании, которые помогают держать аналитику открытой