# Elephant-Goldfish Model: design documents with fresh-eyes review (google/elephant-goldfish-model) — skill card for AI agents

> AgentGid skill card. Checked on 2026-11-13 against commit 8414530 (version 0.1.1).
> Every fact comes from the pack's own files at that commit. Numbers marked "as reported" are the author's.
> Human page: https://agentgid.com/skills/google/elephant-goldfish-model/

> Note: the pack has changed since this check; re-read its files before relying on details.

## Explain it to a non-programmer
Your AI coding assistant helps you think a feature through and writes the design document with you, section by section. Then fresh reviewers who have not seen your chat read it and point out missing assumptions and risks. You decide what to change and what to build.

## Who it is for
- You want a written plan for a feature before your assistant builds it
- You have a design document, README or plan and want it checked against the real code
- You want to find where your documents and your code have drifted apart
- You use Claude Code, Codex, Antigravity or the ChatGPT or Codex desktop app
- Not for: You want a quick fix for a small, well-understood change: the authors say design and review may not pay off there
- Not for: Your assistant cannot start separate helper agents: the review stops instead of running

## What it is
- Kind: command, knowledge
- License: Apache-2.0
- Repository: https://github.com/google/elephant-goldfish-model

## What it does
- Elephant: the main assistant keeps your project, requirements and earlier decisions, and drafts the document
- Goldfish: reviewers start fresh, read only the document and the files it points to, and report gaps
- You: settle trade-offs, approve the design and decide what gets built
- Plain English, no flattery: the drafting skills tell the assistant to stress-test ideas and lead with what is wrong
- Big texts go to files (brief, draft, reports), not into the chat
- When a design choice has more than one sound answer, it shows up to three options and asks you to choose

## Skills
| Skill | Call | Changes code? | In plain words |
|---|---|---|---|
| egm-brainstorm | /egm:egm-brainstorm <idea> (Claude Code, agy CLI); skill picker in Codex | no | Questions you about a rough idea, has fresh reviewers challenge it, and writes a short brief |
| egm-draft | /egm:egm-draft [brief or notes file] | no | Writes the design document with you, one section at a time |
| egm-review | /egm:egm-review <design doc> | no | Has fresh reviewers check a design document in five stages |
| egm-resolve | Runs inside a review; in Claude Code the assistant reads it, it is not a separate command | no | Sorts review findings into fix, contest, ask you, or skip, and edits the document only after your answer |
| egm-review-loop | /egm:egm-review-loop <document> | no | Runs the reviewers you choose against standards you choose, round after round, for any document |
| egm-bootstrap | /egm:egm-bootstrap <folder> | no | Writes or updates a README in each code folder of an existing project, describing what it offers and how to test it |
| egm-drift-report | /egm:egm-drift-report <docs folder> | no | Compares your documents with the current code and lists only where they differ |

### egm-brainstorm
- Asks for: The feature idea in a sentence or two
- Asks for: Answers to open questions: the problem, who uses it, hard limits, edge cases, how success is measured
- Asks for: A go-ahead once the summary of the idea is right
- Gives: concepts_brief.md: overview, analysis, goals and non-goals, alternatives weighed, key decisions
- Gives: 2 to 3 main objections from the reviewers, named in chat
- Gives: A choice: share the brief for team feedback, start drafting, keep refining, or stop
- Say: "/egm:egm-brainstorm Add webhook delivery to this project."

### egm-draft
- Asks for: concepts_brief.md from brainstorm, or your own notes file
- Asks for: Approval or changes after each section
- Asks for: Your choice when a design question has more than one sound answer
- Gives: design_draft.md, section by section: overview and goals, technical plan, alternatives, implementation
- Gives: A final document saved under docs/designs/ (or a folder you confirm)
- Say: "/egm:egm-draft concepts_brief.md"

### egm-review
- Asks for: Path to the design document, README or plan
- Asks for: Optional extra files for the reviewers to read
- Asks for: A review mode: step by step, one combined report, the feature part only, or the implementation plan only
- Gives: zoom_out_report.md, comprehension_report.md, critic_report.md, intermediate_zoom_out_report.md, implementation_readiness_report.md
- Gives: consolidated_review_report.md in the combined mode
- Gives: Findings you can accept, reject or discuss before anything is edited
- Say: "/egm:egm-review docs/designs/webhooks.md"

### egm-resolve
- Asks for: The reviewed document
- Asks for: The list of findings (passed by the review)
- Asks for: Your decision: auto accept, continue, item by item, or your own instruction
- Gives: goldfish_review_report.md with each finding, its severity and a suggested edit
- Gives: Accepted fixes applied to the document
- Say: "interactive breakdown"

### egm-review-loop
- Asks for: The document and the part to review (or the full file)
- Asks for: Which reviewers and which standards
- Asks for: Files or links the reviewers should read
- Gives: Review rounds until two clean rounds in a row, at most 5
- Gives: goldfish_review_report.md after each round
- Gives: A short list of open findings if it stops at the cap
- Say: "/egm:egm-review-loop README.md"

### egm-bootstrap
- Asks for: The folder to document
- Asks for: Your approval if it finds more than 15 folders
- Gives: README.md files (or EGM.md, if you want your own READMEs left alone) with each folder's interface, rules and side effects
- Gives: A top-level README that links the folder READMEs
- Say: "/egm:egm-bootstrap src/"

### egm-drift-report
- Asks for: The folder with design documents or READMEs
- Gives: {date}_DRIFT_REPORT.md: one section per document, the differences or "No drift detected"
- Say: "/egm:egm-drift-report docs/designs/"

## Know before installing
- Not an officially supported Google product, as the README and the plugin listing say, even though it is published under the google organisation.
- Nothing starts by itself: every skill is turned off for automatic use, so you call it by name.
- Install the whole plugin. Copying only the skills folder breaks it, because the skills read the reviewer and standards files.
- It writes files (briefs, drafts, reports, READMEs) into your working folder; you can name another folder.
- A review does not replace tests or your own judgment.
- In Claude Code, use version 2.1.271 or later so reviewers do not see your project instructions file.

## Install (official commands)
- Claude Code: `claude plugin marketplace add google/elephant-goldfish-model` && `claude plugin install egm@egm` — Start a new session after installing. If needed, enable it with claude plugin enable egm@egm. Commands use the /egm: prefix.
- Codex CLI: `codex plugin marketplace add google/elephant-goldfish-model` && `codex plugin add egm@egm` — Start a new session, then pick EGM skills from the skill picker.
- ChatGPT and Codex desktop: Install the published EGM plugin listing in the desktop app (desktop only).
- Antigravity CLI: `agy plugin install https://github.com/google/elephant-goldfish-model` — Keep the /egm: prefix on commands.
- Claude Cowork: Install from claude.ai/plugins or the Cowork plugin panel via the marketplace google/elephant-goldfish-model.
- omp: `omp plugin marketplace add google/elephant-goldfish-model` && `omp plugin install egm@egm` — Skills are called as /skill:egm-<command>.

## What runs on the user's machine
- Code on the machine: no
- Network: optional
- Install: Markdown skills, reviewer prompts, document standards and plugin manifests. No hooks, no scripts in the runtime, no MCP server.
- Reviews: Your assistant starts separate helper agents. In Claude Code they may only read files (Read, Grep, Glob).
- egm-bootstrap: Your assistant runs local commands to list folders and may run a markdown formatter the project already has set up (for example npx prettier), which can reach a package registry.
- Links in a reviewed document: Links that point outside the working folder or to the web are listed and need your approval; approved web pages are saved as files in the working folder for the reviewers.
- Repository only: Development tooling makes network calls (a PR triage script in .github/scripts, the website's star counter and analytics check in site/). It is not part of the skills and never runs from them; the authors leave it out of the runtime release packages.

## Settings
- `EGM artifact directory: <path>`: A line in your prompt, CLAUDE.md or AGENTS.md that sends briefs, drafts and reports to another folder
- `.egmignore`: Folders and files egm-bootstrap skips, like .gitignore

## Examples (from the pack's own files)
- The concepts brief from /egm-brainstorm: output format defined in skills/egm-brainstorm/SKILL.md
- A finding in goldfish_review_report.md: output format defined in skills/egm-resolve/SKILL.md
- The drift report: output format defined in skills/egm-drift-report/SKILL.md

## Steps only the person can do
- In the ChatGPT or Codex desktop app, install the published EGM listing yourself (https://chatgpt.com/plugins/plugins_6aeb90f47ec88191a3c107e01d3eb357) Then: The agent can use the EGM skills in a new local session
- Approve full-tree bootstrapping when egm-bootstrap finds more than 15 folders (skills/egm-bootstrap/SKILL.md) Then: The agent starts parallel workers to write the READMEs
- Choose which links outside the working folder or web pages the reviewers may read (skills/egm-review/SKILL.md (Step 1.3)) Then: The agent saves approved pages as files and passes them to the reviewers
- Approve the design; then ask your assistant to build it (README.md) Then: The agent implements the approved design

## Mistakes to avoid
- Not: Install only the skills folder with a skills installer. Instead: Install the whole plugin for your agent. Why: The skills read the reviewer prompts and document standards at run time
- Not: Run a review in an assistant that cannot start fresh helper agents and expect a result. Instead: Use a host that can start separate agents (Claude Code, Codex, Antigravity CLI). Why: The review skills stop rather than review in the same conversation
- Not: In Claude Code, call egm-resolve or egm-review-loop through the Skill tool. Instead: Start the main skills with /egm:egm-...; inside a review the assistant reads the other SKILL.md and follows it. Why: The skills are hidden from automatic use and the Skill tool refuses them
- Not: Expect /egm-drift-report to bring code and documents back in line. Instead: Read the report, then decide whether to change the document or the code. Why: It only lists differences
- Not: Use an older Claude Code and treat reviews as isolated. Instead: Use Claude Code 2.1.271 or later. Why: Older versions are not verified for reviewer isolation

## What leaves the machine
- Requests for web pages linked from the reviewed document → The linked sites, Only for links you approve during a review. Opt out: Do not approve web links; the review goes on without them
- Package downloads by a project formatter → A package registry, Only if egm-bootstrap runs a markdown formatter your project already configures. Opt out: Deny the command in your assistant's permission prompt

## Lines in the pack aimed at agents (ask the person first)
- "Spawn all Sub-Orchestrators in parallel without" (skills/egm-bootstrap/SKILL.md): On a large tree this starts many helper agents at once (after the user's approval at 15+ folders). Tell the user it can use a lot of tokens and offer a smaller folder first.
- "If a project markdown" (skills/egm-bootstrap/SKILL.md): The skill may run a configured formatter such as npx prettier --write on generated files. Ask before running any npx command.
- "Resolve the brief's `Author` from `git config user.name`," (skills/egm-brainstorm/SKILL.md): The brief records the person's git name. Tell the user before writing it into a file that may be shared.
- "save the raw page of" (skills/egm-review-loop/SKILL.md): Fetching and saving web pages into the working folder writes new files; list them for the user and keep them out of commits unless asked.

## May I use it like this?
- commercial-app: yes (Apache-2.0: keep the license and notices)
- redistribute-files: conditions (Apache-2.0: include a copy of the license and keep notices; no trademark rights)
- client-work: yes (Apache-2.0)

## Which one to pick
- If You want a design document drafted with you and checked by fresh reviewers before anyone builds: skill:google/elephant-goldfish-model
- If You want a full coding workflow that plans, writes tests first and reviews the code it writes: skill:obra/superpowers
- If You want approved specs turned into long autonomous implementation runs: skill:gotalab/cc-sdd

## Works well with
- agent:claude-code: Plugin with read-only reviewer adapters; the authors run end-to-end tests on it
- agent:openai-codex: Codex CLI plugin; reviews need a local executor
- agent:antigravity-cli: Installs with agy plugin install and has named reviewer subagents
- agent:claude-cowork: Installs from the Cowork plugin panel; cloud checks still open

## Works in and needs
- Works in Claude Code: plugin
- Works in Codex CLI: plugin
- Works in ChatGPT and Codex desktop: plugin listing
- Works in Antigravity CLI: plugin
- Works in Claude Cowork: plugin
- Works in omp, Muse: plugin
- Works in Cursor (experimental): plugin
- Works in GitHub Copilot CLI (experimental): plugin
- Needs A host that can start fresh helper agents: Brainstorm and review need separate reviewers
- Needs Claude Code 2.1.271 or later: Reviewer isolation from CLAUDE.md
- Needs Local executor (Codex and ChatGPT): Reviews need it; cloud support is partial

## How to tell it works
Install it, open a project that has a design document or README, and run /egm:egm-review with its path. If the review starts separate reviewers and writes report files with Blocker, Major and Nit findings, it works. If your assistant cannot start helper agents, it stops and says so.

## Safety
Install from github.com/google/elephant-goldfish-model or the published ChatGPT listing. The skills tell your assistant what to do: read them before installing. Give only the files a review needs, since inputs can hold confidential data.
