Google’s spec-first AI coding workflow is now a free Claude Code plugin

Google has published the Elephant-Goldfish Model (EGM) on GitHub, an Apache 2.0 plugin that makes an AI agent draft a design spec and then has fresh, memoryless agents try to break it before any code is written. It installs in Claude Code, Codex and Antigravity.

The release surfaced through a Reddit thread in the Claude community. The details below come from the project’s own repository, google/elephant-goldfish-model.

One long-memory agent drafts, many blank-slate agents attack

The repository describes three roles. The “elephant” is the agent that holds the project context and drafts the specification. The “goldfish” are reviewers that see the draft and its source material but not the earlier conversation, so they cannot share the designer’s hidden assumptions. The third role is the human, who makes the trade-offs and approves the spec before implementation starts.

The plugin ships seven skills:

• Brainstorm: tests an idea and produces a concept brief.
• Draft: turns a brief or notes into a design document.
• Review: checks a design with fresh agents across five stages.
• Review loop: runs a configured group of reviewers on a task.
• Resolve: works through the review findings.
• Bootstrap: creates or updates documentation for an existing codebase.
• Drift report: compares checked-in documents with the current source.

The repository says the project was inspired by an essay by Dave Rensin on Medium about elephants, goldfish and software engineering. It also carries a disclaimer that it is not an officially supported Google product.

Install commands and what it needs to run

For Claude Code, the install is two steps: claude plugin marketplace add google/elephant-goldfish-model, then claude plugin install egm@egm. Codex uses the same pattern with codex plugin, and the Antigravity CLI uses agy plugin install with the repository URL.

The plugin has no fee, according to the repository and an independent walkthrough by the SmartScope blog. The cost is indirect: each review launches several sub-agents, which draw on the host tool’s model quota. SmartScope reported Claude Code 2.1.271 or later as a requirement and recorded a rate-limit interruption in one of its tests. The host must be able to start sub-agents with fresh context. If it cannot, the repository says EGM stops the review rather than falling back to a weaker check.

Why fresh context is the technical bet

This is the technical idea. A language model that wrote a spec in a long session tends to read its own gaps as obvious, because the missing reasoning is still in its context window. Reviewers that start empty must work from the document alone, so anything the document leaves unsaid becomes a visible question. EGM turns that into a procedure rather than a prompting habit.

Two details show how the team treats the limits. First, the isolation is of context, not of the file system. SmartScope notes that reviewers with read access can see any readable project file, so sensitive files should be excluded, and that no security-focused reviewer ships by default, although custom roles can be added. Second, the project states that a review does not replace tests or human judgment, and that the method is meant for features that need design decisions, not small, clear bug fixes. SmartScope also quotes the project’s FAQ as saying that finishing the workflow does not by itself prove higher productivity.

Adoption looks fast. SmartScope recorded 3,463 GitHub stars on October 28, 2026; the repository page now shows about 8.4k stars and 508 forks. Stars measure interest, not results.

The headline claim is Google’s own: the page says “Thousands of Google engineers use EGM.” It cites no measurements of fewer defects, faster delivery or lower cost, and none of it has been independently verified.

The release also fits a pattern. Google recently open-sourced Harness2, which lets a weaker model beat a stronger one by rewriting its own setup. Both projects bet that the scaffolding around a model, not only the model itself, decides how well coding agents perform. EGM differs in keeping a person as the approver at the spec stage.

What is still unproven

Anyone using Claude Code or Codex can try EGM today at no license cost, but the effect on code quality is untested outside Google. The repository does not publish benchmarks, review counts or comparisons with writing a spec by hand, which the Reddit poster says took a full day. Teams weighing it should expect extra quota use and should decide what the goldfish reviewers are allowed to read before running them on a private codebase.