500xEngineer.

All postsSHORT3 min read

Google AI Coding: The Elephant-Goldfish Model

Google's October 22, 2026 post describes an Elephant that designs and fresh Goldfish that review the spec. How to try it in Claude Code and Codex.

Google's AI reviewers have goldfish memory. On purpose. In its October 22, 2026 post, Google described a design session that remembers the project and reviewers that never see your earlier chat. Your spec has to stand alone.

Video not loading? Watch the Short on YouTube

What Google released in October

Google announced the Elephant-Goldfish Model (EGM) plugin on its Open Source Blog on October 22, 2026. It was built internally for Google engineers and released under Apache 2.0.

Google says thousands of Googlers use it every day to ship production Google code. That's Google's own figure. The repo has over 21,000 stars and is still being updated.

The Elephant: design with project memory

The Elephant is a long, context-rich AI session that holds the architectural vision and project memory. It interviews you, debates choices with you and records the decisions. The No Code rule keeps this session focused on design.

The post quotes the line “Design is the new code.” Its stated goal is for the AI to produce the code you wanted the first time. Google gives no speed or bug numbers for that goal.

The Goldfish: review without the earlier chat

Goldfish are stateless review sessions. They examine the draft and referenced source material without the design conversation. Per Google's post, that separation lets them probe for gaps and edge cases without sharing the designer's blind spots.

The design doc has to explain the plan on its own. If a reviewer can't explain it back, the post treats that as ambiguity requiring further drafting.

3 commands to start

  • /egm:egm-brainstorm: load the system context, explore the idea and debate trade-offs under the No Code rule. The agent saves the agreed direction in concepts_brief.md.
  • /egm:egm-draft: turn that brief into design_draft.md. The agent pauses after each section for your feedback. The Markdown doc covers the problem, technical plan, alternatives and an implementation plan with the specific files to change.
  • /egm:egm-review <draft path>: use fresh Goldfish contexts to review the specification.

The 3 review checks in Google's post

Explain it back. A Goldfish reads the complete design doc and explains the architecture and intent back to the Elephant. If it can't accurately summarize the plan, the doc goes back for drafting.

Attack it. Several Goldfish take the role of skeptical technical reviewers, looking for edge cases, hidden dependencies and logical gaps.

Go/no-go. A final Goldfish checks implementation readiness. File paths, data structures and function signatures must be present and unambiguous.

Those are the 3 checks Google's post describes. The repo's review skill lists 5 phases: Initial Zoom Out, Comprehension, Design Critique, Intermediate Zoom Out and Implementation Readiness.

Where it runs and how to install

The October post focused on Antigravity and said other development environments were upcoming. The repo includes install commands for Claude Code and Codex too. Antigravity CLI, Claude Code and Codex pass the repo's automated installation and end-to-end workflow fixtures.

In Claude Code:

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

In Codex:

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

Start a fresh session after installation. In Claude Code and Antigravity CLI, use /egm:egm-brainstorm, /egm:egm-draft and /egm:egm-review <draft path>. In Codex, select the matching EGM skill from the skill picker.

Use an agent that can spawn fresh subagents. Run the brainstorm, draft the doc, then run the Goldfish review. You make the architectural trade-offs and approve the specification before implementation.

The catch

The plugin is free under Apache 2.0. You still need a coding agent that supports fresh subagent contexts. If it can't create those contexts, EGM stops the review.

It's pre-1.0, at v0.1.1. The first release notes say, “This release is not ready for stable-release claims.” The README says, “This is not an officially supported Google product.”

Google gives no speed or bug numbers. The README also says, “A review does not replace tests or human judgment.” Human approval stays part of the process.

The builder angle: my opinion

Sell a design-doc and review pass before a client build. Interview the client with the Elephant, draft the spec and run the Goldfish review. The deliverable is an approved design doc the client signs off on before anyone writes code.

Keep the requirements, trade-offs and review findings in that handoff. This is my opinion, with no income guaranteed.

Daily email

Get each breakdown in your inbox.

One email each evening, right after the Short goes live: the AI tool it covers, the full write-up with sources, and how to get paid for it. Free, and you can unsubscribe anytime.

Read past emails on Substack

Sources

All postsWatch on YouTube