Stress-Test Design Specs with Zero-Context Reviewers Using the Elephant-Goldfish Model

A design doc reviewed by the same agent that wrote it inherits every assumption the conversation baked in. The Elephant-Goldfish Model (EGM), a 3k-star Apache-2.0 plugin published under the google GitHub organization, splits the roles: an Elephant with full project context drafts, Goldfish reviewers with none of that conversation examine the draft, and you decide.

Why This Skill Matters

The README's own framing is the table of three roles. The Elephant holds your project, requirements, and earlier decisions and drafts the specification. Goldfish reviewers see only the document and its referenced source material, without the earlier conversation, which the README says helps expose unstated assumptions and blind spots. You resolve the trade-offs, approve the design, and decide what to build. The README states that thousands of Google engineers use EGM, and its license section adds that this is not an officially supported Google product. The ideas come from Dave Rensin's essay on elephants and goldfish, which the README links.

Use it, in the README's words, when a feature needs design decisions before implementation; for a small, clear bug fix, use your normal coding workflow.

Installation

Pick your coding agent. The README asks you to install the complete plugin, because the skills read the reviewer prompts and document standards; the harness guide is explicit that skill-only installers which copy skills/ alone break EGM, because the skills read agents/ and resources/document_standards/ at run time.

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

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

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

Start a new session after installation. Claude Cowork and ChatGPT Work have their own setup guide in docs/harnesses.md; per that guide, Antigravity, Claude Code, and Codex pass automated installation, reviewer isolation, and end-to-end workflow fixtures, while Claude Cowork and ChatGPT Work have working installation paths with open cloud-trace verification gaps.

Real Workflow: Designing Webhook Delivery

The README's quickstart adds webhook delivery to a project — what should happen when a delivery fails, or the same event arrives twice? The commands below are the Antigravity form; in Claude Code use the /egm: prefix (for example /egm:egm-brainstorm), and in Codex select the EGM skill from the skill picker. Brainstorming and review require a session that can create fresh subagents.

1. Explore the idea. Discuss requirements and trade-offs with the agent; fresh reviewers challenge the proposal, and the agent saves the agreed direction in concepts_brief.md.

/egm-brainstorm Add webhook delivery to this project.

2. Draft the specification. The agent turns the brief into design_draft.md, pausing after each section for your feedback in chat.

/egm-draft

3. Review the draft, using the draft path the agent returned. Fresh reviewers check the design, its clarity, its fit with the code, and its implementation plan. You choose which findings to address.

/egm-review design_draft.md

When you approve the design, ask your coding agent to implement it. The getting-started guide adds two steps the quickstart leaves implicit: send the design to another human as a pull request, because EGM does not replace human oversight, and expect the review to return a lot of feedback by design — take what is relevant and ignore the rest.

The Seven Skills

The README's skill table lists seven, and says you do not have to follow the full sequence:

Bootstrap works bottom-up per the getting-started guide: it generates a README.md for each leaf directory, summarizes upward through parent directories, and compiles the root README into a map of the system. The same guide recommends running the drift report once a week and treating divergence as a human decision — update the doc or the code.

Tips

• Goldfish reviews require fresh subagent contexts; if your agent cannot create them, EGM stops the review rather than running it in the same context.
• To keep briefs, drafts, and reports out of your working directory, the harness guide documents one line for your prompt or your CLAUDE.md / AGENTS.md: EGM artifact directory: .dev/docs.
• The getting-started guide recommends an optional <project>-elephant skill that tells the LLM about your project and links the context you think is most relevant.
• Drafts land under docs/designs/ with a SITEMAP.md index in the recommended layout.

When Not to Use This

The README says it plainly: a review does not replace tests or human judgment, and a small, clear bug fix should go through your normal coding workflow. The whole model depends on a harness that can spawn fresh subagents — without that tool, the review skills stop.

See the leaderboard for more skills.