Skip to content
ML2

Elephant-Goldfish Model: specs reviewed by fresh agents

Added Copyable prompt below

Published Last verified Released

Our verdict

The cleanest packaged version of a pattern that catches agent mistakes before they become code: the session that knows everything writes the spec, and reviewers that never saw the conversation try to break it. It installs in Claude Code, Codex and Antigravity with two commands and works on features that need real design decisions. It is v0.1.0 and Google says it is not an officially supported product, the "thousands of Googlers" usage claim is Google's own with no figure behind it, and every review spawns several fresh reviewer sessions, so skip it for small bug fixes.

Who it's for
Builders and small teams using Claude Code, Codex or Antigravity who keep getting plausible-but-wrong code on features that needed a design first
Leverage
A reviewed, implementation-ready spec from three commands, with gaps found by reviewers that don't share the drafting session's blind spots
Effort
Medium
Date added

What it does

Elephant-Goldfish Model (google/elephant-goldfish-model, Apache 2.0) is a set of agent skills and reviewer personas. The Elephant is your long-context session: egm-brainstorm debates requirements and trade-offs under a no-code rule and saves concepts_brief.md, and egm-draft turns that into design_draft.md section by section, pausing for your feedback. Goldfish are fresh subagents that read only the document and the source it references: egm-review runs a comprehension check (can a fresh agent explain the plan back), antagonistic reviews for edge cases and hidden dependencies, and an implementation-readiness go/no-go. Other skills resolve findings (egm-resolve), document an existing codebase (egm-bootstrap), run a configured reviewer group (egm-review-loop) and compare checked-in docs with current code (egm-drift-report). Reviews need a session that can create fresh subagents; if it can't, EGM stops the review. Claude Cowork and ChatGPT Work have a separate setup guide; Cursor, OpenCode and Copilot CLI support are open issues, not verified yet.

How builders use it

  • Install the full plugin in your agent (commands below) and start a new session so the skills load.
  • Pick a feature that needs design decisions, such as webhook delivery with retries and duplicate events, and run egm-brainstorm on it.
  • Run egm-draft, answer its questions section by section, then run egm-review on the draft and choose which findings to fix.
  • Approve the spec, save it under docs/designs/, then ask your agent to implement it; use egm-drift-report later to see where code and docs diverged.

What to expect

A design document that a zero-context reviewer can understand and implement from, with edge cases and missing details surfaced before the first line of code.

Copy this: Install + first design (Claude Code / Codex)

# Official source: github.com/google/elephant-goldfish-model (Apache 2.0, v0.1.0)

# 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

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

# Start a new session, then (Claude Code / Antigravity use the /egm: prefix; in Codex pick the skill):
/egm:egm-brainstorm Add webhook delivery to this project.
/egm:egm-draft
/egm:egm-review design_draft.md

Related discoveries

See the newest picks →
  • ToolCoding & Shipping Agents

    MCode Stack: MiniMax's coding-agent skills, open-sourced

    MIT set of 17 SKILL.md skills and checklists the MiniMax Code team uses daily, installable as a plugin in Claude Code, Codex or MiniMax Code.

    Effort: LowPrompt#Plugin / extension#Open source

  • ToolCoding & Shipping Agents

    Cloudflare security-audit skill: agent bug hunts

    Cloudflare's MIT skill that turns Claude Code, Codex or Cursor into a six-phase security auditor where a fresh agent tries to disprove every finding.

    Effort: MediumPrompt#Plugin / extension#Open source

  • ToolCoding & Shipping Agents

    claude-mem: memory across sessions for coding agents

    An Apache-2.0 plugin that watches your coding agent's sessions, compresses what it did into searchable notes, and injects the relevant ones into the next session.

    Effort: LowPrompt#Plugin / extension#MCP

  • ToolCoding & Shipping Agents

    improve: shadcn's skill where a strong model plans

    An agent skill (MIT) from shadcn, the creator of shadcn/ui, that uses your most capable model to audit a codebase and write self-contained implementation plans, then hands those plans to cheaper models to carry out. It never edits your source code itself.

    NewEffort: LowPrompt#Plugin / extension#Open source

Get the best new A.I. picks in your inbox

The best new agent workflows and tools we’ve vetted, with a clear verdict on each. Short, occasional, useful.

More in Coding & Shipping Agents →

Free. No spam. Unsubscribe anytime.