Google's Elephant-Goldfish Model plugin reviews AI design docs with fresh agents
Google's open-source Elephant-Goldfish Model plugin has one agent draft a software spec while separate reviewer agents check it without the original conversation.
Google has released Elephant-Goldfish Model, an open-source plugin for coding agents. It separates writing a software design from checking it. One agent, which the project calls the Elephant, develops a specification using the project's requirements and earlier decisions. Separate reviewer agents, called Goldfish, then examine the finished document without the conversation that produced it. They look for missing assumptions and implementation risks.
The separation addresses a specific problem. An agent that discussed a feature with you for an hour still holds that discussion in context, so when it reviews its own draft, missing details may go unnoticed. The article's example is a webhook feature. A fresh reviewer receives only the document and the source files it references, so it has to ask how long retries continue. The reviewer can still read the referenced code. The goal is to remove the unwritten conversation while keeping the evidence needed to judge the design.
The workflow has three stages, each a separate skill. Brainstorming produces a concept brief that records requirements and trade-offs. Drafting produces a design document one section at a time, pausing for your feedback after each section. Review uses fresh reviewers to check whether the document is understandable, whether it fits the codebase, and whether it gives an implementer enough direction. You decide which findings to address and approve the specification before the coding agent starts building.
The webhook example shows what a review is meant to surface. A review might flag several unresolved decisions: whether retries preserve the original event ID, how the receiver detects duplicate deliveries, and what happens when the retry period expires. Each one changes the implementation. Recording the answers before coding gives the agent requirements it can build and test against.
For Claude Code, the install path runs through a marketplace command. You run claude plugin marketplace add google/elephant-goldfish-model, then claude plugin install egm@egm, and start a new session in your project. The commands are /egm:egm-brainstorm, /egm:egm-draft and /egm:egm-review, the last followed by the path to the document. The plugin should be installed whole rather than copied skill by skill, because the skills load reviewer prompts and document standards from the package. Brainstorming and review require an environment that can create fresh subagents. A review run inside the original conversation does not provide the separation the workflow depends on.
The workflow fits features with decisions worth discussing. A small, well-defined fix may not need a specification or several review stages. Fresh reviewers can still miss problems, and tests remain necessary, though they can now target decisions settled before implementation began.
What changed
Google released Elephant-Goldfish Model, an open-source plugin in which one agent drafts a software spec and separate reviewer agents check it without the original conversation.
Type: New capability · Applies to: Elephant-Goldfish Model, Claude Code
Why it matters
Teams using coding agents can surface unwritten decisions, such as how long webhook retries continue, before implementation starts.
What you can do
Try the plugin on your next feature with several open design decisions and confirm that reviews run in fresh subagents.
Original source
We have no link to the original source for this story yet.
Where we saw it
AI Engineering
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works