Google Has AI Write Specs as an Elephant First, Then Pick Holes as a Goldfish

Under Google’s GitHub organization there is an open-source plugin called Elephant-Goldfish Model, or EGM for short. It adds a workflow to AI coding agents: the elephant keeps the project context and drafts the design spec; the goldfish only gets the document and reference materials, does not touch the previous conversation, and focuses on finding unwritten assumptions and blind spots; humans handle architectural trade-offs and approve the spec. The repo’s tagline is blunt: Design with an elephant. Review with a goldfish.

The name comes from Dave Rensin’s article Elephants, Goldfish, and the New Golden Age of Software Engineering. The repo is at google/elephant-goldfish-model, under the Apache 2.0 license. The README states that the project is not an officially supported Google product.

It solves a very specific problem: in the same AI session, an agent first drafts a plan and then reviews it, and it tends to follow along with what it just said. The design point of goldfish review is context isolation. The repo explicitly requires goldfish review to use a new subagent context. If the agent cannot create fresh subagents, EGM stops the review. This limitation looks rigid, but the whole workflow depends on it.

A commenter asked directly in a reply: which model plays the goldfish? Clearing the context with the same model weights may still miss the same things. This question is worth keeping. EGM wraps the process, but which model the reviewer uses and how the prompt is written still determine whether the goldfish truly has fresh eyes.

Installation depends on your coding agent. Claude Code:

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

Codex CLI:

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

Antigravity CLI:

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

ChatGPT and the Codex desktop app can install by opening published plugins in the app. The Antigravity App and IDE extension run /plugin; see the repo docs for details. Start a new session after installation.

The basic flow is not complicated. First run /egm-brainstorm Add webhook delivery to this project., discuss requirements and trade-offs with the agent, fresh reviewers challenge the proposal, and finally the direction is saved to concepts_brief.md. Then run /egm-draft, and the agent turns the brief into design_draft.md, stopping after each section to wait for your feedback. Finally run /egm-review design_draft.md, and fresh reviewers check the design, clarity, fit with the code, and implementation plan. You pick which findings to address, approve the design, and then have the coding agent implement it. In Antigravity and Claude Code, commands need the /egm: prefix, e.g. /egm:egm-brainstorm. Codex selects from the skill picker.

The built-in draft and design-review workflows are aimed at software specs. For other documents, you can use /egm-review-loop to choose reviewers and criteria. Bootstrap and drift report are aimed at software codebases. The repo also provides several standalone skills: Brainstorm, Draft, Resolve, Bootstrap, Review, Review loop, and Drift report. Drift report compares committed documents with the current source code. The demo video is in the repo, with the filename egm-explainer.mp4.

Some commenters said this workflow turns 'forget everything' into work. Others had long been manually opening new chats to poke holes in their own specs and were seeing it packaged as a plugin. Another commenter pointed out that the goldfish is only useful if the spec itself is good enough, so this workflow forces you to write the spec first. That observation is more practical than the name. EGM does not replace testing, nor does it replace human judgment. It turns 'writing it clearly' and 'picking holes in a different context' into fixed actions; the remaining trade-offs are still in human hands.