跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
其他技能

adapt-to-your-project

Run once, right after copying these orchestration files (.claude/skills and .claude/agents) into a project. Reads the project, asks the user only what it can't infer, fills every ADAPT slot under .claude/, removes the sections that don't apply, and writes a short How we work section into the project's CLAUDE.md. Use when the user has just added this starter, or asks to set up, adapt or configure the orchestration for their project.

abidwaqar · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add abidwaqar/AI-Judgment-Orchestration --skill adapt-to-your-project

技能说明


name: adapt-to-your-project description: "Run once, right after copying these orchestration files (.claude/skills and .claude/agents) into a project. Reads the project, asks the user only what it can't infer, fills every ADAPT slot under .claude/, removes the sections that don't apply, and writes a short How we work section into the project's CLAUDE.md. Use when the user has just added this starter, or asks to set up, adapt or configure the orchestration for their project."

Adapt the orchestration to your project

These skills and agents arrive with a marked slot wherever they need a fact about the project:

<!-- ADAPT:<slot-name> — what goes here, and an example -->

In agent frontmatter, where an HTML comment would break the YAML, the same slot is a # ADAPT:<slot-name> — … comment.

Your job is to replace every slot with this project's facts, remove what doesn't apply, and leave the method itself intact. Work in the main conversation, with the user. Here, the user is the person working with you.

The rules while you work

  • Never remove or change an attribution line. Every file carries one: the comment that begins "Adapted from the orchestration behind" (an HTML comment in markdown, a # comment on line 2 of gpt-model.sh). In every file that stays, it stays exactly as it is, in the same place. Count the files that carry it before you start and when you finish.
  • Change facts, not the method. Fill slots and cut what doesn't apply. Don't reword the rules, the round limits, the review recipe or the output contracts. If the project needs a rule changed, say so to the user and change it only on their word.
  • Don't fill this skill's own text. Its slot examples are documentation. Leave .claude/skills/adapt-to-your-project/ out of every search and edit.
  • Edit, don't overwrite. The project's existing CLAUDE.md, AGENTS.md and .gitignore belong to the user: add to them, and never truncate them.
  • Touch nothing else. Edit only .claude/, CLAUDE.md (or AGENTS.md) and .gitignore. Don't install anything, and don't commit: leave both to the user.

1. Read the project

Infer what you can before asking anything:

  • What's here: the README, an existing CLAUDE.md or AGENTS.md (and any area-level ones), and the folder layout two levels deep.
  • The stack: the manifest and lock files (package.json, pyproject.toml, requirements.txt, go.mod, Cargo.toml, Gemfile, pom.xml, build.gradle, pubspec.yaml, *.csproj, composer.json), framework config files, and infrastructure files (Dockerfile, docker-compose.yml, deployment config).
  • The app type: a mobile project (android/, ios/, an app manifest), a website (pages, a dev server, static output), a backend (routes and handlers, no UI), or several.
  • The commands: the scripts in the manifest (build, dev, start, test, lint, typecheck), a Makefile or task runner, and what CI runs (.github/workflows/ or other CI config). CI is often the most honest list of the real checks.
  • The design system: theme or token files, CSS variables, a Tailwind or similar config, a components folder, font files, a logo.
  • The voice and names: the product's own copy, in its UI strings and its README.
  • Git and GitHub: git remote -v; the default branch (git symbolic-ref refs/remotes/origin/HEAD); long-lived branches such as develop or staging; whether gh is installed and logged in (gh auth status), and whether the repository has issues (gh repo view --json hasIssuesEnabled).
  • Codex: codex --version and codex login status. If both work, independent reviews can run.
  • Seeing the app: which tools this session has for it, such as a mobile MCP, a browser MCP (Playwright) or Claude Code's built-in browser, and whether an emulator, a simulator or a local URL is already set up.
  • Every slot: grep -rn "ADAPT:" .claude --exclude-dir=adapt-to-your-project. Note each file, line and slot name. The slots are stack, app-type, see-the-app, design-tokens, voice, requirements-home, commands, domain-expert and branch-model.
  • The attribution lines: grep -rlF "Adapted from the orchestration behind" .claude | wc -l. Note the count.

2. Ask only what you couldn't infer

A few questions at a time, two or three per message. For each, say what you inferred and from where, and give your recommendation, so most answers can be a "yes".

  1. The app type: a mobile app, a website, a backend, or a mix, and the sizes and themes it supports.

  2. How the AI should see the running app:

    • an emulator or simulator plus a mobile MCP, for a mobile app;
    • a browser tool, such as Playwright MCP or Claude Code's built-in browser, for a website;
    • or none, for a backend, where the engineer checks the running service through its API and logs instead.

    If the tool they pick isn't installed, say which one and how to add it.

  3. The design tokens and voice: where they live, or "none yet". With none yet, the rules stay, and the ui-ux-designer skill's §6 No tokens yet applies: the designer proposes provisional tokens alongside the first item's mockups, the user approves them with that item's experience, the requirements carry them, and that item's delivery builds them into the code.

  4. Codex, if the check in step 1 failed: will they install it (npm install -g @openai/codex, then codex login) for independent reviews? If not, say plainly what that means: every review then runs on the delivery-reviewer agent (and every UX review on the ui-ux-designer agent), Claude reviewing Claude's work, and every verdict is recorded as NOT INDEPENDENT. It's their choice; record it in step 5.

  5. Where requirements live: GitHub issues (the default), or file mode: one markdown file per item in `.claude/requirements

阅读完整来源 ↗