Guide · Agent skill
How to use the Goldfish Review Loop agent skill (egm-review-loop)
Updated · written with AI from the project's own documentation and checked against our catalogue data
The egm-review-loop skill lets an AI agent audit a design doc, README, feature spec or implementation plan with several separate reviewer subagents, then evaluate their findings and apply the fixes you approve. This guide explains what the skill does, how to install it, how to trigger it with plain requests, and what it needs before you start. By the end you will be able to run a structured review round on one of your documents.
Goldfish Review Loopgoogle/elephant-goldfish-model · Apache-2.0 Download the skill Quick start
- 1Download and unzip. Download the egm-review-loop ZIP from this page and unzip it into ~/.claude/skills/ or your project's .claude/skills/ folder.
- 2Confirm the EGM files. Make sure the Elephant-Goldfish Model files (agents/, resources/document_standards/ and docs/) are available to the agent.
- 3Prepare your target. Choose the document and, if needed, the section to review, and back it up or commit it to version control.
- 4Ask for the review. Make a natural request naming the file and the review type, for example a feature spec or implementation plan.
- 5Confirm the configuration. Check the target, reviewers and mapped standards the agent displays, then confirm or correct them.
- 6Decide on findings. Read goldfish_review_report.md and choose Auto accept, Continue or Interactive breakdown at the decision gate.
What the skill makes the agent do
The skill runs the Antagonistic-Review process from the Elephant-Goldfish Model (EGM). The main agent, called the Elephant, keeps the working context. Reviewers, called Goldfish, are subagents started in fresh, isolated contexts with no conversation history, so they catch unstated assumptions instead of inheriting the author's assumptions.
Each round, the agent re-reads the target section, has the reviewers audit it against written standards, consolidates their findings with severity levels (Blocker, Major, Nit), and asks you to decide what to do. Approved text repairs are applied to the target file. The loop then repeats in rounds.
Install the egm-review-loop skill
Download the skill ZIP from this page and unzip it. For a personal install, place the folder in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder. Alternatively, in the Claude apps, open Settings, go to Capabilities, and upload the skill under Skills.
The skill's files are agents/openai.yaml and SKILL.md. The skill points to other files, including agents/, resources/document_standards/ and docs/, which must be available to the agent. The reviewed material says these were not included in the listing, so confirm they are present in your Elephant-Goldfish Model repository.
- Personal: ~/.claude/skills/
- Project: .claude/skills/ inside your project
- Claude apps: Settings → Capabilities → Skills → upload
Trigger the skill with a natural request
The skill is set to disable-model-invocation: true, so it is meant to be invoked deliberately rather than picked up automatically. Name the document and the kind of review you want. The skill's description matches requests about reviewing design docs, READMEs, feature specifications, implementation plans, and checking completeness and tone.
If you omit details, the agent will prompt you for the missing inputs one at a time: the target file, the semantic target (a section such as a header ID, or full file), the reviewers, the standards, and a reading list of context files. It then shows the plan and asks you to confirm before reviewing.
- "Run the Goldfish review loop on docs/payments-feature-spec.md, full file."
- "Review the implementation plan in plans/search-rewrite.md with the readiness and codebase details reviewers."
- "Check the README for completeness and tone using the EGM review loop."
Understand the inputs and reviewer choices
The skill takes five inputs: TargetFile, SemanticTarget, Reviewers, Standards and GoldfishContext. Reviewer shortnames such as design_critic, structural, style, implementation_readiness, codebase_details, comprehension and zoom_out map to files in agents/. Standard shortnames such as feature-spec-standard and implementation-readiness-standard map to resources/document_standards/. A name containing a slash is read as a custom path.
For feature specs, the skill recommends design_critic and structural. For implementation plans, it recommends implementation_readiness and codebase_details.
Review the findings and decide on fixes
Before editing, the agent categorises findings as CONTEST or PROPOSE FIX, saves a goldfish_review_report.md file, and presents a decision gate with the options Auto accept, Continue and Interactive breakdown. Only approved repairs are written to the target file.
Findings you contest, or record as accepted risk, are settled and are not reopened in later rounds. An applied fix does not settle a finding, because a later reviewer may raise the same issue again.
Safety notes: scripts and network
According to the listing, this skill runs no scripts and uses no network access. Its work happens through the agent's subagent tool, file reading, the report file and file editing. The loop can change your target file after you confirm, so keep your document under version control or back it up before a run.
The skill also says that if your harness has no subagent tool, the agent should stop. A review in the same context is not a Goldfish review, so do not expect a valid result from a single-agent run.
- Requires a subagent tool: Claude Code Agent, Codex spawn_agent, or Antigravity invoke_subagent
- Edits apply to the target file only after you approve them
- Slow and token-heavy, so it is not suited to quick proofreading
What people use it for
- Checking a design doc for unstated assumptions before a review meeting
- Testing an implementation plan for readiness and missing codebase details before work starts
- Auditing a README for completeness and tone before publishing
Frequently asked questions
How do I install the egm-review-loop skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use or a project's .claude/skills/ folder for one project. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.
What does the egm-review-loop skill require?
It requires an agent harness with a subagent tool, such as Claude Code Agent, Codex spawn_agent or Antigravity invoke_subagent. The Elephant-Goldfish Model repository files, including agents/, resources/document_standards/ and docs/, must also be available to the agent.
Is the egm-review-loop skill free?
The skill is licensed under Apache-2.0. The source material does not state pricing for the skill or for the agent harness you run it in, so check those separately.
Does the skill run scripts or use the network?
According to the listing, it runs no scripts and uses no network access. It does edit your target file after you approve fixes, so back up the document first.
Why does the skill need a subagent tool?
Reviewers must run in fresh contexts with no conversation history to catch unstated assumptions. The skill states that a review in the same context is not a valid Goldfish review, so it stops if no subagent tool is available.
When should I pick a different tool instead?
For quick one-off proofreading, a simpler check is better because the multi-round subagent loop is slow and token-heavy. Also avoid it when you do not want an agent editing important documents.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use this skill if you want a structured, multi-reviewer second opinion on design docs, feature specs, implementation plans or READMEs before a review meeting or build starts. Skip it for quick proofreading, for documents you do not want an agent to edit, or in an environment without a subagent tool.
Download the skill