Google Elephant-Goldfish Plugin Guide (2026): Install EGM in Claude Code, Codex and Antigravity

Google has open-sourced the Elephant-Goldfish Model (EGM) plugin, a design-first workflow for AI coding agents. One agent with full context writes the plan, then fresh agents with zero context try to break it before any code is written.

It installs in Claude Code, Codex and Google Antigravity. Here is what it does, how to install it and how to use each command.

What the Elephant-Goldfish Model Is

EGM is a method first built inside Google for its own engineers. Google’s open source blog says thousands of Googlers use the plugin every day to ship production code. The public release came on October 22, 2026.

The idea is summed up by co-author Dave Rensin: “Design is the new code.” You spend your effort on a solid written spec, and the coding agent builds from it.

• The Elephant: a deep-context session that holds the architecture, history and your intent. It brainstorms with you and drafts the spec, but writes no code.
• The Goldfish: stateless subagents with no memory of your chat. They read only the spec and the live source code, so they share none of the Elephant’s blind spots.
• The human: you own the architecture, the trade-offs and the final approval.

Note: the README says this is “not an officially supported Google product”. It is open source under Apache 2.0, with no paid tier.

Why Fresh-Context Reviewers Help

A long chat builds up shared assumptions. When the same agent reviews its own plan, it fills gaps from memory and misses what a newcomer would trip on.

A Goldfish has none of that memory. If it cannot understand the spec, the spec is unclear. If it finds an edge case, the Elephant never covered it.

Google’s blog describes three Goldfish checks:

1. Comprehension check: a Goldfish explains the design back. If it gets it wrong, the doc is ambiguous.
2. Antagonistic review: several Goldfish act as sceptical reviewers hunting for edge cases, hidden dependencies and logic gaps.
3. Implementation readiness: a final Goldfish gives a go or no-go on file paths, data structures and function signatures.

The current plugin expands this into stages called Zoom Out, Comprehension, Critic and Readiness, run by named personas such as Design Critic, Naive Reader, Tech Lead and Codebase Auditor.

How to Install the Plugin

Install the whole plugin, not just the skills folder. The skills read persona files and document standards at run time, and skill-only copies break.

Claude Code install, step by step

1. Update Claude Code. Make sure your version supports plugins. My Claude Code install guide covers updates.
2. Add the marketplace. Run claude plugin marketplace add google/elephant-goldfish-model.
3. Install the plugin. Run claude plugin install egm@egm.
4. Restart your session. Start a new Claude Code session in your project folder.
5. Check the commands. Type /egm: and you should see the EGM skills, such as /egm:egm-brainstorm.

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

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

Tip: new to plugin marketplaces? My Claude marketplace guide explains how they work and how to remove one.

All Seven EGM Commands Explained

In Claude Code, prefix each command with egm:, for example /egm:egm-draft. In other harnesses use the plain name.

Step by Step: Your First Feature With EGM

Here is the core loop from Google’s getting started guide, using a simple example feature such as “add CSV export to reports”.

1. Brainstorm. Run /egm-brainstorm and describe the idea. The Elephant interviews you under a “no code” rule, and a Goldfish swarm stress-tests the concept.
2. Draft the spec. Run /egm-draft. It writes docs/designs/<feature>.md section by section: problem, technical plan, alternatives and a file-level implementation plan.
3. Review with Goldfish. Run /egm-review docs/designs/<feature>.md. Fresh subagents check clarity, gaps and readiness.
4. Resolve findings. Run /egm-resolve to accept or reject each finding. Treat feedback as a menu, not orders.
5. Get human review. Open a pull request with just the design doc so a teammate can approve it.
6. Build it. Ask your coding agent to implement the approved spec. Small mistakes get point fixes. Big misses mean the spec needs another pass.
7. Check for drift. Run /egm-drift-report later to see where code and spec have split.

Your project ends up with a docs/designs/ folder holding a SITEMAP.md index and one Markdown spec per feature. These docs become memory for future agents and new teammates.

Using EGM on an Existing Codebase

Goldfish only know what is written down. On an old project with thin docs, they will flag everything as unclear.

• Bootstrap docs: run /egm-bootstrap path/to/project to write READMEs from leaf folders up to the root. Review them yourself.
• Add an Elephant skill: optionally create a small project skill (the guide suggests a name like <project>-elephant) that describes your project and links key code and docs.
• Schedule drift checks: the guide suggests running /egm-drift-report weekly as a recurring task.

If you already keep an AGENTS.md file, it pairs well with this. See my AGENTS.md guide for how agents read project instructions.

Harness Support and Subagent Limits

Goldfish reviews need a truly fresh context. If a harness has no subagent tool, the review skills stop rather than run in the same context.

Gemini CLI, Cursor and OpenCode are not listed in the official harness docs. The “ask any agent to install it” route may work, but it is not tested by the project. For more on Claude’s subagents, see my Claude Code subagents guide.

When to Use It (and When Not To)

Google’s own guide warns against forcing the design-first process on simple bug fixes. Use it where a bad plan is expensive.

• Good fit: new features, cross-module changes, data model changes and anything a teammate must review.
• Good fit: long agent runs, where a vague spec leads to hours of wrong code.
• Poor fit: typo fixes, one-line bugs and quick experiments.
• Cost note: swarms and multi-stage reviews mean several model calls. The docs give no token figures, so watch usage on paid plans.

Warning: each Goldfish is a separate subagent, so a full review can burn through usage limits faster than a normal chat. Check your plan limits before running big review loops.

My Take

EGM puts a name on something good engineers already do: write the plan, then hand it to someone who was not in the room. Automating that “someone” with zero-context agents is clever and cheap.

It pairs naturally with plan modes in coding tools, like Claude Code plan mode. Plan mode keeps the agent from coding too early. EGM adds a hostile review before you say yes.

Troubleshooting: Common Problems and Fixes

• Commands not found in Claude Code: use the /egm: prefix and start a new session after installing.
• Review stops with no output: your harness may not expose a subagent tool. EGM refuses to review in the same context by design. Switch to a supported harness.
• Skills fail to find files: you probably copied only the skills/ folder. Reinstall the full plugin so agents/ and the document standards are present.
• Goldfish flag everything as unclear: your codebase lacks docs. Run /egm-bootstrap first.
• Nesting errors: if a harness cannot nest subagents, use direct dispatch instead of the synthesizer, as the harness docs advise.
• Hitting usage limits: run /egm-review on smaller specs, or split a big feature into several design docs.

Frequently Asked Questions

Is the Elephant-Goldfish plugin free?

Yes. It is open source under Apache 2.0. You only pay for the model usage of the agent you run it in.

Does it work with Claude Code?

Yes. Install it with claude plugin marketplace add google/elephant-goldfish-model and claude plugin install egm@egm, then use the /egm: commands.

Does it work with Gemini CLI or Cursor?

They are not in the official harness list. The tested harnesses are Antigravity, Claude Code and Codex, with Claude Cowork and ChatGPT Work also installable.

Does the Elephant write code?

No. The Elephant only designs and drafts the spec. Your coding agent builds the code after you approve the design.

Where are the design docs saved?

In docs/designs/ in your project, with a SITEMAP.md index and one Markdown file per feature.

Is it an official Google product?

It comes from Google engineers and the Google open source blog, but the README says it is not an officially supported Google product.

When was it released?

Google announced the open source plugin on its open source blog on October 22, 2026.

Sources

• Google Open Source Blog: Introducing the Elephant-Goldfish Model plugin
• GitHub: google/elephant-goldfish-model README
• EGM docs: Getting started
• EGM docs: Harnesses
• egm.dev: Elephant-Goldfish Model site