AboutAI ToolsCursorClaude CodeLovableBolt.newAI tool reviewsTutorialsVibe codingAI agentsMobile apps with AIAll tutorialsBlogSponsorWork With MeJoin 285K+ Subscribers

8 November 2026 · 3 min read

Google Open-Sources Elephant-Goldfish Model Plugin for Spec-First AI Coding

Google released the Elephant-Goldfish Model plugin under Apache 2.0, a workflow that forces AI coding agents to draft a specification and have it reviewed by fresh, zero-context subagents before any code is written.

On 8 November 2026, Google open-sourced the Elephant-Goldfish Model (EGM) plugin, a design-and-review layer intended to make AI coding agents write a specification before they touch code [1]. The plugin is not a new language model but a set of skills, reviewer personas and document standards under the Apache 2.0 licence. Google says thousands of its own engineers already use the workflow internally.

What the plugin does

EGM tackles two common pain points: AI “slop” from long conversation contexts and the difficulty of catching wrong assumptions early. The core idea comes from Dave Rensin’s essay on elephants and goldfish [1]. An Elephant is a long-running agent session that holds project context and drafts a design document. Goldfish are fresh, stateless subagents that review the draft with no memory of previous conversations. Only when the spec passes review does implementation begin.

Three slash-commands drive the workflow:

  • /egm-brainstorm clarifies the idea and then runs a six-lens Goldfish swarm. The lenses cover YAGNI, token cost, UX, migration, observability and tech debt. The output is a concepts_brief.md file.
  • /egm-draft turns the brief into a structured design document under docs/designs/, written section by section.
  • /egm-review launches the five-stage Goldfish review.

A separate /egm-drift-report command later compares the implemented code with the original spec. The documentation recommends running it weekly.

The five-stage Goldfish review

The review skill runs five sequential stages. Each stage may trigger at most one re-audit, and all findings are ranked as Blocker, Major or Nit [1]:

  1. Initial Zoom Out – high-level check of the entire document.
  2. Comprehension Test – a Goldfish paraphrases the design. If the paraphrase is wrong, the spec is deemed ambiguous.
  3. Critic Review – design, structural and style reviewers hunt for gaps and edge cases.
  4. Intermediate Zoom Out – rechecks architecture after fixes.
  5. Implementation Readiness – validates file paths, data structures and function signatures.

If the host agent does not support spawning fresh subagents, EGM stops. A same-context review is explicitly not a Goldfish review.

Supported coding agents

At launch the harness guide covered Antigravity CLI and App, Claude Code (version 2.1.271 or later), Claude Cowork, Codex, ChatGPT, omp and Muse [1]. Antigravity CLI, Claude Code and Codex pass automated end-to-end test fixtures. For Claude Code, reviewers receive only Read, Grep and Glob tools, and CLAUDE.md files are excluded from their context.

Concrete test data from the source:

  • Codex 0.163.0 completed 15 native fresh-context reviewers in order; two live tests passed in 573.13 seconds.
  • Codex CLI with the v0.1.0 ZIP passed the full suite in 482.30 seconds.
  • A Claude Cowork run stopped after stages 5 to 7 because of a usage limit.
  • An open issue (#237) notes that AGENTS.md files leak into Codex reviewers, indicating isolation is not yet fully guaranteed on every host.

EGM is a plugin with no model weights. It uses the host’s existing model quota, so the only cost is the token consumption of extra review passes. No accuracy benchmark has been published; the evidence is internal adoption and harness test runs [1].

How EGM compares with similar tools

The source article includes a comparison with GitHub Spec Kit, Kiro Specs and the Superpowers plugin [1]. The key difference is that EGM reviews the specification itself with zero-context agents before any code exists. Spec Kit and Kiro centre on human review between phases. Superpowers isolates subagents mainly at implementation time, not at the specification stage.

What it means for developers

For teams building with AI coding agents, EGM adds a lightweight gate that catches misunderstandings while they are still cheap to fix — in a Markdown file rather than in a pull request. The forced separation between the spec-writing Elephant and the reviewing Goldfish brings a form of independent review that is otherwise hard to maintain in long agent sessions. Developers should be aware of the extra token usage and verify that their chosen harness truly isolates Goldfish contexts, as the leakage reports suggest some platforms still need hardening. The plugin is free to use and extend, so teams can adapt the personas and lenses to their own conventions.

Sources

  1. Google AI Releases Elephant-Goldfish Model Plugin for Spec-First Coding
  • elephant-goldfish
  • coding-agents
  • spec-first
  • google
  • open-source

Researched and drafted with AI from the sources above.

Learn to build with AI

Hands-on tutorials on Cursor, Claude Code, AI agents and more.