Your Design Reviewer Shouldn't Remember Your Design Meeting

You've probably been here: you spend an afternoon working through a design with your AI assistant, filling it in on all the context and constraints, and it produces a spec you're happy with. Then you ask it to review that spec and it just... agrees with you. Of course it does. It was there when you wrote it. The Elephant-Goldfish Model is an open-source plugin built around a simple fix for that problem: let one model hold the context and draft the spec, then have separate reviewers look at the result without any of the earlier conversation.

What It Does

EGM is a plugin for AI coding assistants that splits design work into two roles. The "elephant" holds your project context and drafts the specification. The "goldfish" reviewers then examine that draft (along with any referenced source material) without the earlier conversation, so they come at it cold. You make the architectural trade-offs and approve the spec before anyone implements it.

The built-in draft-and-review workflow is aimed at software specifications, though the README notes you can swap in different reviewers and standards with /egm-review-loop for other kinds of documents. Bootstrap and drift reporting are specific to software codebases. It installs across ChatGPT and Codex desktop, Antigravity, Claude Code, and Codex CLI, with setup notes for Claude Cowork and ChatGPT Work.

Why It's Cool

Most AI review tools assume more context is better. EGM bets the opposite, and that's the whole idea.

• Fresh eyes are the feature. The reviewers don't see the conversation that produced the draft, which is exactly the condition that tends to surface unstated assumptions and blind spots. If you've ever had a reviewer nod along because they'd already absorbed your reasoning, you'll get why this matters.

• You stay in the loop. The model drafts and the reviewers critique, but the README is clear that you make the architectural trade-offs and approve the spec. It's a division of labor, not an autopilot.

• It works on documents you already have. You don't need to write a new spec to get value. /egm:egm-review docs/designs/webhooks.md reviews an existing design doc, README, or PR plan saved as Markdown against the live code, checking the design itself, its clarity, its fit with the codebase, and the implementation plan.

• Drift reporting is a genuinely useful extra. Point /egm:egm-drift-report at a folder and the agent compares each document with the current source code and reports where they've diverged. You decide whether to update the doc or the code. Anyone who's maintained design docs knows how quickly they rot.

• It's a plugin, not a platform. It slots into the agent you already use, and the install steps are one or two commands.

How to Try It

Pick your agent and install the full plugin (the skills need it to read the reviewer prompts and document standards).

For Claude Code:

claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egm

For Codex CLI:

codex plugin marketplace add google/elephant-goldfish-model
codex plugin add egm@egm

For Antigravity CLI:

agy plugin install https://github.com/google/elephant-goldfish-model

ChatGPT and Codex desktop users can install the published plugin from the desktop app (that listing is desktop-only). Start a new session after installing.

If you want a quick test, run the review on a document you already have:

/egm:egm-review docs/designs/webhooks.md

or check a whole folder for drift:

/egm:egm-drift-report docs/designs/

A couple of practical notes from the README: the /egm: prefix applies in Antigravity (agy CLI) and Claude Code, and you should keep it on agy CLI 1.4.8, where unprefixed names don't get completed. In Codex, you select the matching EGM skill from the skill picker instead. Brainstorming and review also require a session that can create fresh reviewers, so check that your setup allows it.

The repo is at github.com/google/elephant-goldfish-model, with fuller setup guides in docs/harnesses.md and docs/getting_started.md.

Final Thoughts

EGM is a narrow, opinionated tool, and that's fine. It's for people who already use an AI assistant for design work and have noticed that the same assistant makes a poor critic of its own output. The fresh-reviewer idea is easy to explain and easy to test on a document you've already written, which makes it a low-risk thing to try. If you maintain design docs alongside code, the drift report alone might earn the install.