All postsLONG-FORM5 min read
Elephant-Goldfish Model: I Tested Google's Plugin in Claude Code
One unattended Claude Code run tested Google's EGM planning plugin: 774 lines of design, fresh reviewers, real findings and a caption fix we applied by hand.
Google announced its Elephant-Goldfish Model plugin on October 22, 2026. We tested it in one overnight run with nobody at the keyboard: an agent answered from a brief we wrote, and the session covered planning and review only.
The short answer
It installed and ran in Claude Code. Fresh reviewers found real design problems we could check against our files. The plugin wrote a 774-line plan for a change to a 35-line script, with no implementation. My take: useful scrutiny, substantial overhead for this small fix.
What Google released
EGM supplies prompts and reviewer roles inside a coding assistant; the assistant's model does the work. Google released the plugin under Apache 2.0. Google says thousands of Googlers use it daily to ship production code. The repository also says it is “not an officially supported Google product.”
The October 22 post describes 3 Goldfish checks and an Antigravity-first goal. The repository we installed from on November 18, 2026 describes 5 review stages and installation paths for Claude Code, Codex and Antigravity.
Elephant and Goldfish
The Elephant is a long session that remembers the conversation and writes the design. Goldfish are fresh reviewers without that conversation history, reading the plan and referenced files. A reader walking in cold can expose assumptions the author remembers but never wrote down.
The person approves the plan. Drafting presents defensible designs and trade-offs for the author to choose.
Install it in Claude Code
Use Claude Code 2.1.271 or later. Run:
claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egm
Start a fresh session afterward. The core commands are:
/egm:egm-brainstormwritesconcepts_brief.md./egm:egm-draftwritesdesign_draft.md, pausing for section approvals./egm:egm-reviewwrites stage reports; uninterrupted full review also writesconsolidated_review_report.md.
For a document you already have, point review at a design, README or saved plan: /egm:egm-review docs/designs/webhooks.md.
Our run
Conditions:
- The two install commands ran against a throwaway config folder (plugin version 0.1.1), so our real setup stayed unchanged. Each session then loaded that installed copy.
- Headless Claude Code 2.1.327.
claude-opus-5-5for the main session and reviewers.- One run, with answers and approvals from a written brief.
- A Mac under load.
We used a sanitized slice of our video pipeline. Episode 23 had two real release-night problems: a caption rule joined “Astra is” into “AstrAIs,” and YouTube Studio refused a description containing 6 arrows with angle brackets. Both had been fixed by hand. The requested feature was a release text lint that reports problems without rewriting copy; repairing the caption rule was explicitly deferred.
| Phase | Turns | Session time | Reviewer runs | What was written |
|---|---|---|---|---|
| Brainstorm | 4 | 448 s (7 min 28 s) | 7 plus 1 synthesizer | concepts_brief.md: 332 lines, 2,824 words |
| Draft | 5 | 370 s (6 min 10 s) | 0 | design_draft.md: 774 lines, 5,798 words |
| Review | 1 | 706 s (11 min 46 s) | 10 | 5 stage reports and consolidated report: 7,981 words |
Total session time was 25 min 24 s, with 17 reviewer runs plus 1 synthesizer. No code was written. Claude Code reported 6,879,082 tokens, including 5,834,199 cache reads, 798,946 cache writes, 245,685 output and 252 uncached input. Most were rereads. This ran on a subscription.
Brainstorm caught an import trap: reusing the caption writer's rules by importing it would execute its top-level write and overwrite hand-fixed captions. We confirmed the missing main guard and write. We also checked the draft's word rule: 1 problem in reconstructed faulty captions, 0 in hand-fixed captions, and 6 angle brackets in the description. The faulty file had not been retained.
What the reviewers found
Reviewer-reported stage counts include duplicates.
| Stage | Reviewers | Blocker | Major | Nit |
|---|---|---|---|---|
| 5 Initial Zoom Out | zoom-out, inventive-catalyst | 0 | 3 | 5 |
| 6 Comprehension | comprehension | 0 | 2 | 8 |
| 7 Critic | design, structural, style | 0 | 13 | 25 |
| 8 Intermediate Zoom Out | zoom-out | 0 | 4 | 4 |
| 9 Implementation Readiness | implementation-readiness, codebase-details, style | 0 | 10 | 23 |
That totals 97 raw findings: 0 blockers, 32 major and 65 nits. Consolidation merged them into 12 major actions and 6 groups of nits. The verdict was “Not ready to implement.”
Three findings mattered most. First, the design preserved the known caption bug while adding a check around it. The rule matches a word ending in “a” followed by one starting with “i.” We checked our earlier scripts: 5 of 27 contained matching pairs, 12 pairs altogether. These are scripts at risk; both words must land in the same caption for corruption to occur.
Second, the design called a false alarm its worst outcome, then proposed blocking releases without an override, supported by a hand calculation on 1 episode. Three stages raised this contradiction. The remedy proposed was to try older episodes before enabling blocking.
Third, the design promised no raw error traceback while handling only decode errors. Permission failures could still crash it.
Reviewers also needed checking. One tentatively called vm a potentially unused allowlist entry. The Elephant contested it; we checked both pattern-use locations and confirmed the original patterns still run. The Elephant contested 5 findings overall. Stage 8 found that 3 proposed fixes from stages 6 and 7 would worsen stage 5 findings.
Nits included “6” versus “six,” uneven horizontal rules and the author name sounding like marketing. Some were functional, including an empty captions file passing. My recommendation: read major findings first and expect to skip plenty of nits.
What we changed afterwards
We applied the word-boundary fix by hand in the build that captioned this video, so matching cannot begin inside a word. The plugin did not write it. On Episode 23's hand-fixed captions, the old rule changes 1 line; the corrected rule changes 0. The lint it designed was not built.
Limits in EGM's own docs
The guide says feedback is abundant by design: choose what matters. Reviews do not replace tests or human judgment. It recommends skipping design-first for a simple fix.
Claude reviewers start with clean conversations and read-only tools. This provides context and tool isolation; reviewers can still read files accessible to the session.
What we did not test
We did not implement the design, repeat the run, change models or harnesses, or test a live human conversation. Resolve, lite, bootstrap, drift reporting and release ZIPs were also untested. These results establish no improvement in finished code.
How I'd use this: my opinion
I'd review an existing plan with real choices, keep a person approving it, and verify important findings against source. Fresh reviewers earned their place here by challenging the preserved cause and contradictory rollout. For our small fix, the paper stack was overkill.
What's the last plan an AI wrote for you that you'd want a reviewer with no memory to read?
Disclosure: narration is AI-generated with ElevenLabs using a clone of my own voice. Opinions are mine.
Daily email
Get each breakdown in your inbox.
One email each evening, right after the Short goes live: the AI tool it covers, the full write-up with sources, and how to get paid for it. Free, and you can unsubscribe anytime.
Sources
- Google's October 22 announcement
- November 18 repository snapshot: README, Claude Code guide, workflow guide, draft skill, review skill.
- Our recorded November 18 lab run, source checks and subsequent hand-applied caption fix.