What Is Google's Elephant-Goldfish Plugin? Review the Plan Before the Code

Hey hey hey, welcome to Agentic Pod! I'm Liam.

And I'm Stella! Liam, Google just open-sourced the plugin its own engineers use to catch AI coding mistakes. And the reviewer is a goldfish?

A goldfish on purpose. The repo is called elephant-goldfish-model, and the tagline is "Design with an elephant. Review with a goldfish."

Okay, um, I need that explained. Why would you want a reviewer that forgets?

So, the one who wrote the plan can't see its own blind spots. By the end you'll know how it works, how to install it in Claude Code or Codex, and when it's not worth the trouble.

Start simple. What is it, in one sentence?

It's an open-source plugin that makes your coding agent get its design reviewed before it writes any code. Apache 2.0 license, and the README says it's not an officially supported Google product.

So no promises from Google. Got it. Who are the animals?

The elephant is your main session. It holds all the project context, and it drafts the specification. The goldfish are fresh reviewer sessions that never saw your conversation.

So they only see the draft and the files it points to.

Exactly. Nothing else. If the draft says "use the usual auth flow" and the goldfish has no idea what that means, that's a finding. An unstated assumption just got caught.

Oh, that's, like, clever. It's like asking a stranger to read your instructions.

Right, the stranger test. And you stay the boss. You make the trade-offs and approve the spec before anyone writes code.

Okay, but how does that actually run? Is it one command?

It's seven slash commands. In Claude Code and Antigravity they start with slash egm colon. First, egm-brainstorm tests an idea and writes a concept brief.

And then?

egm-draft turns that brief into a design document. Then egm-review sends fresh agents to read the draft, pointed at your design draft.

And the findings?

egm-resolve walks you through them, so you agree on each change. There's also egm-bootstrap, which writes docs for an existing codebase, and egm-drift-report, which compares your checked-in docs to the current source.

Ooh, uh, drift report. So it tells me the docs lied.

Politely. And egm-review-loop runs a configured group of reviewers for one task. So the order is brainstorm, draft, review, resolve, then you build.

And where do I get it? Because I'm lazy.

Three ways, straight from the README. In Claude Code, run claude plugin marketplace add google slash elephant-goldfish-model. Then claude plugin install egm at egm.

And Codex?

Same shape. codex plugin marketplace add the same repo, then codex plugin add egm at egm. In Antigravity it's agy plugin install, followed by the GitHub URL.

And then I just start typing slash commands?

Start a new session first. The README says so. Then try a small feature you've been putting off. Write the idea, run egm-draft, then egm-review, and read what the goldfish complains about.

And I should do that before the agent writes anything.

Yes, that's the whole point. A wrong plan is cheap to fix in a document and expensive to fix in a pull request. And that brings us to a catch.

There's always a catch. What is it?

First, the goldfish needs fresh sessions. If your agent can't spin up fresh subagent contexts, the README says EGM stops the review. It won't pretend.

Good. A fake review would be worse than none.

Second, it doesn't replace tests or human judgment. The README says that itself. It reviews the design, not whether your code runs.

And one analysis I read says there's no security reviewer by default.

Right, so, you know, don't treat it as a security audit. The personas are customizable, so you can add one. Third, a small change doesn't need a whole review ceremony.

So not for renaming a variable.

Please no. Who's it for, then? Anyone whose agent writes big features from a loose idea.

Google says thousands of its engineers use it every day to ship production code. That's their claim, not a measured result, but it tells you where the idea came from.

And why does it matter outside Google?

Yeah, because agents write code fast now, so the weak spot moved. It's the plan. A bad assumption in a draft gets built perfectly, and that's expensive.

My verdict: try it on one real feature this week, and keep your hand on the approve button.

Same. The repo, the Google blog post and our step-by-step guide are in the description. Would you let a goldfish review your next plan? Tell us in the comments. See you next time!