--- title: "grounded-engineering" url: "https://claudemods.dev/zh-tw/builds/grounded-engineering-d6299a" source: "https://github.com/madjagstudios/grounded-engineering/tree/main/plugin" source_type: "github" repo: "https://github.com/madjagstudios/grounded-engineering/tree/main/plugin" author: "madjagstudios" author_handle: "madjagstudios" stars: 2 category: "工作流程與 CI" categories: ["workflow"] tags: ["agents", "engineering-practices", "hooks", "pane", "plugin", "skill-repos"] permission_level: 2 scan_complete: true install: "claude plugin marketplace add madjagstudios/grounded-engineering\nclaude plugin install grounded-engineering@grounded-engineering" added: "2026-10-09T00:06:28.285Z" updated: "2026-10-08T21:21:52Z" locale: "zh-tw" content_locale: "zh-tw" --- # grounded-engineering 附有來源依據的工程實務，以及經過審查的技能儲存庫，顯示在 Claude Code 面板中。 ## Description 工程實務取自成熟的開源儲存庫，每條都附上它的來源；另外還有一批經過審查的第三方技能儲存庫。它以模組形式在 Claude Code 內執行，也能以外掛安裝為即時面板，並提供一個命令列工具。輸入 /grounded，面板會讀取你開啟的儲存庫，顯示適合的實務（例如：有測試卻沒有 CI 關卡、只有 Claude 會讀取的代理指示、沒有作業系統沙箱），並附上理由與固定的來源連結。 Adapt 會請 Claude 提出最小的變更建議；在你核准之前，不會寫入任何內容。 此外也列出十六個第三方技能儲存庫，它們都已在固定的 commit 上審閱過，並提供「Explain」安裝流程。另有一個 adopt 命令列工具，可把受管理的區塊寫入 AGENTS.md、CLAUDE.md 或中性的 Markdown 檔案。 ## Install ``` claude plugin marketplace add madjagstudios/grounded-engineering claude plugin install grounded-engineering@grounded-engineering ``` ## Original README # Grounded Engineering [![Validate](https://github.com/madjagstudios/grounded-engineering/actions/workflows/validate.yml/badge.svg)](https://github.com/madjagstudios/grounded-engineering/actions/workflows/validate.yml) [![npm](https://img.shields.io/npm/v/grounded-engineering?label=npm)](https://www.npmjs.com/package/grounded-engineering) [![License: MIT](https://img.shields.io/github/license/madjagstudios/grounded-engineering)](LICENSE) [![Node](https://img.shields.io/node/v/grounded-engineering)](package.json) [![Claude Code mod](https://img.shields.io/badge/Claude_Code_mod-2.1.288%2B-D97757)](#claude-code-mod) Engineering practices drawn from mature open-source repositories, each tied to the source it came from, plus a shelf of reviewed third-party skill repositories. It runs as a mod inside Claude Code, a live pane installed as a plugin, and as a command-line tool. ![The Grounded pane in Claude Code: it lists the practices that fit the open repository, and Adapt asks Claude to propose the change](docs/media/grounded-pane.gif) Type `/grounded` and the pane reads the repository you have open. It shows the practices that fit what it finds (tests with no CI gate, agent instructions only Claude reads, no OS sandbox) and why each one applies. **Adapt** asks Claude to propose the smallest change that brings the practice into your repository; nothing is written until you approve it. Every card links to the sources it was drawn from, pinned to the revision that was read. ## Install ### Claude Code mod Needs Claude Code 2.1.288 or later, in the terminal or the desktop app's Code tab. Mods install as plugins: ```text /plugin marketplace add madjagstudios/grounded-engineering /plugin install grounded-engineering@grounded-engineering ``` Then type `/grounded`. In the terminal the pane also has keyboard shortcuts: `1` and `2` switch screens, `d` shows what it detected, and `a` runs the open card's main action. ### Command-line tool Needs Node.js 20 or later. It writes the same practices into `AGENTS.md`, `CLAUDE.md` or a neutral Markdown file; see [Adopt a profile](#adopt-a-profile). ## Skill repos ![The Skill repos screen: Ponytail is featured at the top, and Explain install has Claude read the repository before anything is installed](docs/media/grounded-skill-repos.gif) The second screen lists sixteen third-party skill repositories, each read at a pinned commit and approved by a maintainer before it is listed. The shelf holds links and our own short notes only; credit and stars go to the authors. **Explain install** has Claude read the repository at that commit and list what it would add (hooks, scripts, network use, settings it changes) before quoting the author's own install steps. Claude installs nothing unless you agree, and you type any slash commands yourself. Authors who ask to be removed are delisted. The records live in [`research/skill-repos/`](research/skill-repos/). ## Repository layout ```text research/ Source observations, pinned references, and category audits practices/ Short, reusable engineering-practice cards integrations/ Consumer-specific translation guidance for agent instruction files plugin/ The Claude Code mod: pane, skills, and generated catalog src/, bin/ The command-line tool scripts/ Validation, catalog build, and source-drift tools ``` ## Adopt a profile Adoption does not rewrite existing policy. `preview` only reads, and `create` stores a proposal under `.grounded-engineering/proposals//` for you to review; nothing is written until `apply --confirm`. ```bash # one-off, nothing installed npx grounded-engineering adopt preview --profile ai-assisted --adapter claude # or install the command npm install -g grounded-engineering grounded-engineering adopt create --profile ai-assisted --adapter claude # Review proposal.yaml, plan.md, and diff.patch; complete local_decisions. grounded-engineering adopt apply --confirm grounded-engineering check ``` Profiles: - `baseline`: the original eight cards on repository context, code quality, testing and verification. Its pack metadata stays at `v0.2.0` so existing adopters stay green. - `ai-assisted`: all seventeen cards. Adapters: - `neutral` (default) writes Markdown to `docs/grounded-engineering.md` if the repository has a `docs/` folder, otherwise to `GROUNDED_ENGINEERING.md`. - `codex` writes to `AGENTS.md`. If `AGENTS.override.md` exists, Codex reads it instead, so the tool reports it and writes nothing. - `claude` writes to the repository-root `CLAUDE.md`. It reports `.claude/CLAUDE.md`, nested `CLAUDE.md` and `CLAUDE.local.md` files but does not edit them. Every adapter writes only inside managed blocks keyed by card ID, and leaves everything outside them byte for byte. Apply writes the target and `.grounded-engineering/manifest.yaml` only after re-checking the proposal's preconditions. A repository can have one adapter target. Adding a second target, and choosing cards with `--cards` outside `preview`, are not supported yet. `grounded-engineering check` reads the manifest and the target files and compares them with the pack bundled in the installed CLI. It exits `0` when clean, `1` on drift or a repository-state mismatch, and `2` on an invocation error. The write path is specified in the [adopt apply policy](policies/adopt-apply.md); compatibility notes for each release are in [`CHANGELOG.md`](CHANGELOG.md). ## How sources are used Cards paraphrase their sources and link to the revision and the file or heading they draw on. We quote only where a source's license allows it, and never copy vendor prompt files. Each card says whether its practice is observed, recommended, or validated in use. A card does not change when its source does. `npm run check:sources` reports drift, and a card marked validated records which revision of each source it was checked against. ## Contributing Start with [`CONTRIBUTING.md`](CONTRIBUTING.md) and [`research/README.md`](research/README.md). Keep source observations in `research/`, the practice in `practices/`, and tool-specific wording in `integrations/`; a change that mixes them is harder to review. Grounded Engineering is built and maintained with AI agents. ## License Original repository content is available under the [MIT License](LICENSE). Third-party sources remain under their own licenses and terms; links and attribution are recorded with the evidence.