Elephant-Goldfish Model

Elephant-Goldfish Model (EGM) is an open-source plugin for AI coding assistants, published by Google under the Apache 2.0 license and noted as not an officially supported Google product. It splits spec work into two roles: a context-rich Elephant agent that explores the problem and drafts the specification, and Goldfish reviewers that critique the draft without the earlier conversation. The approach is introduced in an essay by Dave Rensin.

What It Is

EGM is a set of skills installed into coding agents such as Claude Code, Codex, and Antigravity. The Elephant holds project context, requirements, and prior decisions to brainstorm and write a design document. Goldfish subagents are stateless and isolated, reading only the document and its referenced source material, to surface missing assumptions, unclear requirements, and implementation risks. The human makes the architectural trade-offs and approves the spec before implementation. The built-in draft and design-review workflow targets software specifications.

Workflow and Skills

The documented path runs from an idea to a verified spec. /egm-brainstorm tests an idea and writes a concept brief (concepts_brief.md). /egm-draft turns a brief or existing notes into a design document, pausing after each section for feedback. /egm-review has fresh agents review the design across five stages, and /egm-resolve works through findings and agrees on changes. Once the spec is approved, the coding agent builds it, and /egm-drift-report compares checked-in documents with the current source code. Other skills are /egm-bootstrap, which creates or updates documentation for an existing codebase, and /egm-review-loop, which runs a configured group of reviewers for a specific task, including non-software documents. Each skill can be used on its own.

Setup and Limits

Installation is through agent plugin commands for Antigravity CLI, Claude Code, and Codex CLI, plus a published desktop plugin listing for ChatGPT and Codex. Brainstorming and review need a session that can create fresh subagents; if the agent cannot, EGM stops the review. The README states that a review does not replace tests or human judgment.