GitHub radar
Review a feature's design before your AI agent writes code
This is a plugin for your AI coding agent (Claude Code, Codex, ChatGPT): it adds a discussion-and-independent-review step for the spec before any code gets written, catching gaps and forgotten edge cases early.
One 'elephant' session holds your project context and drafts a plan (a spec document). Separate 'goldfish' sessions review that plan with fresh eyes, knowing nothing of the earlier conversation, to find blind spots.
When it helps
Useful before building a non-trivial feature, when edge cases and architecture decisions matter. Skip it for small fixes or simple tasks — it's unnecessary overhead there.
Pros
- Catches gaps and forgotten edge cases before any code is written
- Runs on top of your existing agent — Claude Code, Codex, ChatGPT
- No GPU or separate model install needed
- You always make the final call and approve the plan
Cons
- Needs an agent that can spawn separate fresh subagent sessions
- You pay in your regular agent's tokens — usage goes up
- Built for software specs; other documents need manual setup
How to set it up — step by step
- 1Open your project in Claude Code and run: claude plugin marketplace add google/elephant-goldfish-model
- 2Install the plugin itself with: claude plugin install egm@egm
- 3Start a new session — the review won't work without one
- 4Send the agent the text from the block below, filling in your own feature
- 5Read and discuss the draft plan (design_draft.md) the agent saves after review
- 6Approve the plan — only then ask the agent to write the code
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install the google/elephant-goldfish-model plugin for me in Claude Code: add the marketplace and install egm@egm, then run /egm-brainstorm for my feature idea, walk me through the draft and review, and show me the final design_draft.md. Ask me if you need any permissions.
No GPU or separate model install needed — it's a plugin that runs through the AI coding agent you already use (Claude Code, Codex, ChatGPT), billed through that agent's usual token cost.
Open on GitHub▌ More finds