bama
BAMA: 모든 Mac 앱을 만들 수 있습니다
AI 코딩 에이전트로 사양 기반 개발을 위한 오픈소스 방법·스킬셋, 깊은 컨텍스트 에이전트가 설계 초안, 제로 컨텍스트가 코드 전 검토한다
An open-source method and skill set for spec-driven development with AI coding agents. Deep-context agents draft the design. Zero-context agents review it before any code is written.
스킬PythonGitHub @kurcontko온라인운영자 미확인
써 보셨어요? 한 번만 눌러 주세요. 로그인은 필요 없어요.
아직 감지된 업데이트가 없습니다. GitHub 릴리스가 나오면 여기에 보입니다.
개인 계정의 공개 저장소 소유자는 GitHub로 로그인해 운영자를 확인하고 소개를 직접 관리할 수 있습니다. 관리자 승인은 필요하지 않습니다.
로그인한 뒤 확인 파일을 만들어 이 저장소의 기본 브랜치에 30분 안에 커밋해 주세요. 다른 브랜치나 포크의 파일은 인정하지 않습니다. 파일을 올린 뒤 아래에서 확인하면 관리자 승인 없이 관리할 수 있습니다. 자격을 유지하려면 파일을 남겨 두세요.
로그인이 필요하면 로그인 화면으로 이동합니다. 로그인 후 이곳으로 돌아와 원하는 확인 버튼을 다시 눌러 주세요.
프로젝트 폴더에서 /nomorevibe verify를 실행하면 소유권을 확인하고 소개를 직접 관리할 수 있습니다.
1분 만에 확인하기Claude Code plugin for Google's Elephant-Goldfish Model (EGM): spec-driven development with isolated Goldfish reviews
저장소 README에서 가져옴
README 전체를 불러오는 중입니다.
README 전문 보기 ↗A Claude Code plugin for the Elephant-Goldfish Model (EGM), Google's open-source method for spec-driven development with AI coding agents.
An Elephant (your long-running session) brainstorms and drafts a design doc. Goldfish reviewers (fresh subagents with no conversation history) then check whether the doc stands on its own before any code is written. The skills, personas, and document standards come unchanged from google/elephant-goldfish-model. This repository adds only what Claude Code needs to run them correctly.
Community project. Not an official Google or Anthropic product. Upstream officially supports Antigravity and Gemini CLI. Native Claude Code support is proposed upstream in google/elephant-goldfish-model#101.
Requires Claude Code 2.1.271 or later.
/plugin marketplace add kurcontko/egm-claude-code
/plugin install egm@egm-community
Or from a shell:
claude plugin marketplace add kurcontko/egm-claude-code
claude plugin install egm@egm-community
| Step | Command | Who |
|---|---|---|
| 1. Clarify an idea into a concepts brief | /egm:egm-brainstorm <idea> | Elephant |
2. Draft the design doc in docs/designs/ | /egm:egm-draft | Elephant |
| 3. Review it with isolated reviewers | /egm:egm-review docs/designs/<doc>.md | Goldfish |
| 4. Work through the findings | /egm:egm-resolve | Elephant |
| 5. Build from the reviewed spec | your usual prompt | |
| Weekly: compare docs against code | /egm:egm-drift-report | |
| Existing codebase: generate READMEs bottom-up | /egm:egm-bootstrap |
/egm:egm-review-loop runs an automated review-and-fix loop over any document.
The skills run only when you invoke them; Claude never starts them on its own.
See egm.dev and the
upstream getting-started guide
for the method itself.
Upstream's skills are harness-neutral: they say "spawn a subagent" and
reference agents/ and resources/ by relative path.
scripts/sync_upstream.py copies upstream at the
commit pinned in upstream.json and applies a small,
mechanical overlay:
| Change | Why |
|---|---|
Reviewer personas get tools: Read, Grep, Glob and omitClaudeMd: true | Without these, Claude Code subagents can edit files and load your CLAUDE.md, so a Goldfish is not really context-free |
The synthesizer and bootstrap orchestrator get a note with the plugin root and the egm: agent names; the synthesizer is limited to Agent, Read, Grep, Glob, Write | They run as subagents, so they never see a skill's note |
Each skill gets a note with the resolved ${CLAUDE_PLUGIN_ROOT} and a pointer to docs/claude-code.md | Installed skills live in the plugin cache, so agents/ and resources/ would otherwise resolve against your project. docs/designs/ stays in your project |
docs/claude-code.md maps the skills' wording onto Claude Code tools | invoke_subagent becomes Agent with egm:egm-<persona>, forks are ruled out because a fork inherits the whole conversation, and chained skills are read from the plugin because the Skill tool refuses them |
Skill folders are renamed to their frontmatter names (egm_review → egm-review) | Claude Code names plugin commands after the folder |
Everything else is copied byte for byte. These edits are safe here because
this plugin only targets Claude Code. Native support proposed upstream in
#101 keeps
skills/ and agents/ untouched instead, because Antigravity rejects
tools: Read, Grep, Glob in a shared persona file.
python scripts/sync_upstream.py --check rebuilds the plugin from the pinned
commit and fails if plugins/egm/ differs.pytest tests checks the overlay: reviewer frontmatter, skill notes, resource
paths, and that the manifest matches the marketplace and lock file.tests/live_probe.py runs one headless Claude Code session (about $0.15)
in a throwaway workspace. It confirms that /egm:egm-review reads the guide
from the plugin root and spawns egm:egm-zoom-out as a new agent, not a fork.
It also confirms that the reviewer sees neither a CLAUDE.md marker nor a
parent-prompt marker, and that it has no write, shell, or agent tools.CI runs the first two plus claude plugin validate. A weekly workflow rebuilds
from upstream main and opens a pull request when upstream changes.
Edit overlay/ or scripts/, never plugins/egm/ directly, then run:
python scripts/sync_upstream.py # rebuild from the pinned commit
python scripts/sync_upstream.py --ref main # or move the pin to upstream main
pytest tests
Apache-2.0, the same as upstream. See LICENSE and NOTICE for what is changed from the upstream files.