Neodrop
Create
FeedDiscoverMy ChannelsAuto-publishJoin Our Discord
Neodrop
A smoother experience in the app
FeedDiscoverCreateMy Channels
google/elephant-goldfish-model: 5.5K stars this week for five review stages that start without your chat
2026-11-15

google/elephant-goldfish-model: 5.5K stars this week for five review stages that start without your chat

This week's pick is Google's Elephant-Goldfish Model plugin: an elephant session drafts your design document and five stages of zero-context reviewers attack it before any code is written — with per-host install steps, the reviewers' real reception, and the two hosts where the goldfish still sees your instruction files.

This Week's Trending Agent Skills
This Week's Trending Agent Skills@NeoDrop Official
google/elephant-goldfish-model is a plugin of skills and reviewer personas, not a model. An "elephant" session holds your project context, interviews you and drafts a Markdown design document — and writes no code. Then "goldfish" subagents, spawned with no memory of that conversation, read the document cold and try to break it. You approve the design; only then does an agent implement it. 12
GitHub's weekly Trending board put it 9th when the board was read on 15 November 2026, with 5,539 stars added inside the seven-day window. 3 The repository carries about 19,600 stars and 1,456 forks under the Apache 2.0 licence, with 72 open issues and 39 open pull requests counted separately. 145 Google announced the plugin on 22 October 2026; the repository itself dates from 13 September, and the last push landed at 5:33 a.m. Pacific on 15 November. 12
The plugin manifest reports version 0.1.1. 6 The skills.sh directory counts 3.6K installs. 7 Google's own announcement says the plugin "was developed internally for Google software engineers" and that "thousands of Googlers use this plugin every day to ship production Google code" — a claim the company has not quantified. 2
Loading content card…

What ships inside the plugin

Three kinds of file do different jobs in a session, and the counts are small enough to check.
PieceCountWhat it changes in a session 1
Skills7The commands: /egm-brainstorm, /egm-draft, /egm-review, /egm-resolve, /egm-review-loop, /egm-bootstrap, /egm-drift-report. 8
Reviewer personas8, plus two more filesNamed subagents with their own frontmatter: a skeptical architect, a hostile design critic, a naive reader, a pragmatic tech lead, a structural validator, a copy editor, a codebase auditor, and a TRIZ contradiction-hunter. The other two files are a synthesizer that de-duplicates the swarm's findings and a bootstrap orchestrator. 9
Document standards5Rubrics a reviewer applies: a feature-spec standard, the Heilmeier standard, an implementation-readiness standard, a re-implementation standard, and a zoom-out standard. 10
The Heilmeier standard is the interesting one in a coding plugin. It is the catechism a program manager applies before funding work — what is being attempted, how it is done today, what is new, who cares, what the risks and payoffs are, what it costs, how long it takes, and what the mid-term and final exams are — and here it is applied to a design document before an agent is allowed to build from it. 10
The no-code rule is not a preference. In the elephant phase, the session debates requirements and trade-offs and writes concepts_brief.md and then design_draft.md; code comes after your approval. 8
Google's EGM lifecycle diagram
Google's own lifecycle diagram, published with the plugin on 22 October 2026: brainstorm with the elephant, draft with the elephant, then three goldfish checks on the finished document. Image: Google Open Source Blog. 2

The five stages of a review

/egm-review is not one prompt with a stern tone. It is five stages, numbered 5 to 9, each spawning fresh reviewers, each writing its own report file, each pausing for you at a checkpoint. 11
StageWhat it checksPersonas spawnedReport file 11
5, Initial Zoom OutThe whole document at altitude: is the shape right, does it reuse what existsegm-zoom-out, egm-inventive-catalyst, through the synthesizerzoom_out_report.md
6, Comprehension TestWhether a reader with no prior knowledge can restate the designegm-comprehensioncomprehension_report.md
7, Critic ReviewArchitecture, document structure, and prose, under the feature-spec standardegm-design, egm-structural, egm-stylecritic_report.md
8, Intermediate Zoom OutWhether the architecture still holds after the fixes from stages 6 and 7egm-zoom-out plus the synthesizerintermediate_zoom_out_report.md
9, Implementation ReadinessFile paths, data structures and signatures — everything needed to build without deciding anythingegm-implementation-readiness, egm-codebase-details, egm-styleimplementation_readiness_report.md
Stage 6 keeps the reviewer's paraphrase in the report, not just its findings, because a paraphrase that drifts is itself the finding: the document failed to explain itself. 11 Findings are ranked Blocker, Major or Nit, and each stage allows one re-audit — after that the skill tells you the cap is reached rather than looping. 11
Two rules carry more weight than the rest. If your harness cannot spawn a subagent, the skill stops instead of reviewing in the same context, because "a review in the same context is not a Goldfish review". And the persona prompts ban flattery outright: no "great", "brilliant" or "spot-on" for your ideas, no apologies when corrected, critique before merit. 11
There are four ways to run it: interactive with a checkpoint after each stage, consolidated into one report, feature-spec only, or implementation-plan only. 11

Where the fresh context comes from, and where it leaks

The skills do not name a vendor's tool. They say "spawn a subagent" and name a persona; each host maps that to its own mechanism — invoke_subagent in Antigravity, the Agent tool in Claude Code, native spawn tools in Codex — and where a host has no named personas, a general-purpose subagent is told to read the persona file first. 12
That mapping is where the method's central promise gets weaker. Issue #237, open since 30 October 2026, reports that on Codex, reviewers receive the project AGENTS.md and the user-level ~/.codex/AGENTS.md even with fork_turns: "none", and that on Antigravity they receive the project AGENTS.md and GEMINI.md. On Claude Code the adapters set omitClaudeMd: true and a control run returned UNKNOWN for a marker planted in every instruction file. 13
The issue's own argument for why that matters is better than any summary: "these files often hold the team's assumptions about the system ('we always use X', 'Y is out of scope'). A reviewer that starts with them is reviewing with the author's priors, which is the bias the Goldfish step is meant to remove." 13
The repository's guidance already concedes the gap in general terms: a fresh conversation "does not mean that the host omits its instruction files", because fork_turns: "none" isolates parent turns rather than project instructions. 12 What the issue adds is that the same guide lists Antigravity CLI, Claude Code and Codex as passing reviewer isolation, when what passed in the recorded tests was parent-turn isolation. 1314
The practical fix is yours to make: put the assumptions you want challenged into the review prompt as explicit criteria, and keep only stable workspace conventions in AGENTS.md. 12

Which agents it reaches

HostInstall routeWhat has actually been verified 12
Claude Codeclaude plugin marketplace add google/elephant-goldfish-model then claude plugin install egm@egmAutomated install and end-to-end fixtures pass; needs 2.1.271 or later; reviewers get only Read, Grep and Glob. 1415
Antigravity CLI, App, IDE extensionagy plugin install https://github.com/google/elephant-goldfish-modelFixtures pass on CLI 1.4.5 with 16 native reviewers; the standalone Antigravity IDE is not supported because it has no fresh-context subagent tool. 14
Codex CLI and desktopcodex plugin marketplace add google/elephant-goldfish-model then codex plugin add egm@egmFixtures pass; a live suite on 0.163.0 completed 15 native reviewers in 573.13s. No named personas; a general-purpose subagent reads the persona file. 14
Claude CoworkMarketplace google/elephant-goldfish-model from claude.ai/pluginsInstall works; the recorded run stopped at stage 8 on a usage limit, and isolation evidence is an open gap. 16
ChatGPT desktop and WorkThe published EGM plugin listing, or a workspace import of .agents/pluginsNine native reviewers completed; the fresh-context probe returned UNKNOWN and cloud traces remain an open gap. 17
ompomp plugin marketplace add then omp plugin install egm@egmCommunity-verified on v18.12.6: all seven skills resolve and all reviewer personas register; no end-to-end fixture run. 18
MuseLocal-path install, then muse plugins approve egmFull fixture on 1.4.14 with 15 fresh reviewers; tool restrictions do not apply there, so reviewers are not read-only. 14
Windsurf, now DevinWorkspace routing files, experimentalA Devin CLI run on 10 November passed routing, read-only reviewers and isolation, but "no workflow ran to a final artifact"; the desktop IDE is untested. 19
CursorA .cursor-plugin/ directory shipsAn experimental adapter was merged; verified support is still open, with native run, permission enforcement and fresh-context verification outstanding. 20
ClineNothing in the repositoryNot documented, and not in the install table. 12
Gemini CLINothing in the repositoryGoogle's own route here is Antigravity; GEMINI.md appears in the docs only as a file that leaks into Antigravity reviewers. 13
If you work in Cursor or Cline, this is not your week yet, and the reason is specific rather than vague: the goldfish stage needs a host that can spawn a genuinely fresh subagent with a read-only toolset, and Cursor's adapter has not been verified against that bar. 20

Install it

  1. Claude Code. claude plugin marketplace add google/elephant-goldfish-model then claude plugin install egm@egm. 12
  2. Antigravity CLI. agy plugin install https://github.com/google/elephant-goldfish-model. On the App or the IDE extension, run /plugin or install into ~/.gemini/config/plugins/egm/. 12
  3. Codex CLI. codex plugin marketplace add google/elephant-goldfish-model then codex plugin add egm@egm; the desktop app takes the published listing. 12
  4. Start a new session, then call the skills with the host's prefix: /egm:egm-review in Claude Code and Antigravity, the skill picker in Codex, @ in ChatGPT. Keep the prefix on older Antigravity CLI builds, where an unprefixed name is not completed. 2122
Install the whole plugin, not just the skills. A skill-only installer copies skills/ alone, and the run breaks, because the skills read agents/ and resources/document_standards/ while they work. 12
Briefs, drafts and reports land in your working directory unless you say otherwise. A line in your prompt, or in CLAUDE.md or AGENTS.md — EGM artifact directory: .dev/docs — moves them; a path that resolves outside the working directory needs confirmation, and in a headless run the files fall back rather than guess. 12

First test: sixty seconds on a document you already have

You do not have to adopt the whole lifecycle to see what the reviewers do. Point the review at something you wrote already — a design note, a PR plan saved as Markdown, a README:
/egm:egm-review docs/designs/webhooks.md
Fresh reviewers check the design, its clarity, its fit with the code, and its implementation plan. 21 Then ask the drift check to compare a folder of documents against the current source:
/egm:egm-drift-report docs/designs/
The agent reports where each document and the code have diverged, and you choose which side to change. The docs suggest running it weekly. 23
If the output earns it, the full loop is three commands — /egm-brainstorm to a concepts_brief.md, /egm-draft to a design under docs/designs/, /egm-review to the five stages — with a human review of the pull request in between, which the docs call essential rather than optional. 8
Budget for many model calls. The repository's own test record counts 15 native reviewers for a single review on Codex, 16 on Antigravity, a live Codex suite at 573.13 seconds, and a Claude Cowork run that a usage limit interrupted at stage 8. 14 Two more limits from the same record: egm-bootstrap was "larger than the load limit" on the Antigravity fixture, and the project's own guidance says not to force the design-first process onto a one-line bug fix. 814

What users actually said

The largest discussion is a Reddit thread from 31 October 2026, 804 points at 92% upvoted and 112 comments, written by someone who had spent a day per project writing a spec by hand before prompting. 24 His account of why: killing a bad sync architecture on paper saved weeks of code.
The replies are mostly not impressed. "Congrats on reinventing openspec!" runs one. "Oh we're doing animal metaphors now? That's a nice level of misdirection and required prior knowledge, just enough to feel novel - ship it." Another is two words: "Overengineering at it's finest." 24
The thread's auto-generated summary says the consensus was a "collective eye-roll, but a helpful one", that readers saw well-established planning practice renamed, and that one of the EGM leads agreed in the comments that they were reintroducing spec-driven development for the AI age. 24 The fair version of that objection is hard to answer: anyone who writes a plan file and then opens a fresh chat for implementation already runs a rough elephant-and-goldfish loop, and the plugin's contribution is orchestration, personas and report files rather than the idea.
Reach ran the other way. The official @GoogleCloudTech post on 7 November had 2,156 likes and 200,231 views. 25 The highest independent number belongs to @thisguyknowsai on 31 October: 3,441 likes, 253,447 views, 6,431 bookmarks. 26 A Google product manager's explainer on 8 November reached 30,135 views. 27 None of the posts fetched for this piece describes a finished project that came out of an EGM run; they describe the method.
There is no accuracy benchmark either. The repository's evidence is adoption and recorded harness runs, and MarkTechPost's review of the launch says the same thing directly: "no accuracy benchmark published". 28

Adopt it, or keep your own spec habit

A CLAUDE.md, a plan file and a fresh chat for implementation already get you the shape of this method. EGM adds three things that setup does not: named reviewers with different lenses, a fixed five-stage order with report files you can hand to another person, and a drift check that keeps the documents honest after the code moves. 1123
Four conditions make it the wrong tool:
  • You work in Cursor or Cline today. Verified Cursor support is an open issue; Cline appears nowhere in the repository. 20
  • Your token budget is tight. One review spawns 15 or 16 reviewers, and a single recorded run hit a usage limit mid-pipeline. 14
  • Nobody will maintain docs/designs/. Approved specifications are the artifact; an abandoned folder turns the drift report into a chore that reports your own neglect. 8
  • You want proof before belief. There is no measurement of whether EGM-reviewed designs ship better code. 28
The status matters as much as the method: Apache 2.0, version 0.1.1, 39 pull requests open, and "This is not an officially supported Google product" in the repository's own words. 5621
If you do one thing this week, run the review against a document you already trust, and read what comes back before you let it near your next feature.

Quick reference

FieldDetails
Pickgoogle/elephant-goldfish-model, an Apache-2.0 plugin of skills and reviewer personas. 1
Best forMulti-file features where a wrong assumption is expensive, and teams that will keep design documents current. 8
HostsClaude Code 2.1.271+, Antigravity CLI and App, Codex, Claude Cowork, ChatGPT, omp, Muse; Cursor experimental; Cline and Gemini CLI unsupported. 12
InstallWhole-plugin install per host; skill-only installs break the run. 12
NeedsA host that can spawn fresh subagents, and plan headroom for 15–16 reviewers per review. 14
LicenceApache 2.0; not an officially supported Google product. 21
Version0.1.1, with 39 pull requests open and a push on 15 November 2026. 56
AdoptionRank 9 on the weekly board with 5,539 stars in the window; about 19,600 stars, 1,456 forks, 3.6K directory installs. 37
First test/egm:egm-review <existing doc>, then /egm:egm-drift-report <design folder>. 21
Main caveatsInstruction files leak into Codex and Antigravity reviewers (#237); no accuracy benchmark; Cursor unverified; reviews are many model calls. 13

References

  1. 1
    Repository

    github.com

  2. 2
    Google Open Source Blog

    opensource.googleblog.com

  3. 3
    GitHub Trending, weekly window

    github.com

  4. 4
    Open issues

    github.com

  5. 5
    Open pull requests

    github.com

  6. 6
    plugin.json

    github.com

  7. 7
    skills.sh

    skills.sh

  8. 8
    docs/getting_started.md\

    github.com

  9. 9
    agents/\

    github.com

  10. 10
    resources/document_standards/\

    github.com

  11. 11
    skills/egm-review/SKILL.md

    github.com

  12. 12
    docs/harnesses.md

    github.com

  13. 13
    Issue #237

    github.com

  14. 14
    TESTING.md

    github.com

  15. 15
    docs/claude.md\

    github.com

  16. 16
    Issue #105\

    github.com

  17. 17
    Issue #106\

    github.com

  18. 18
    Issue #186\

    github.com

  19. 19
    Issue #376\

    github.com

  20. 20
    Issue #145\

    github.com

  21. 21
    README

    github.com

  22. 22
    Issue #220

    github.com

  23. 23
    skills/egm-drift-report/SKILL.md

    github.com

  24. 24
    Reddit thread

    reddit.com

  25. 25
    @GoogleCloudTech on X

    x.com

  26. 26
    @thisguyknowsai on X

    x.com

  27. 27
    @Saboo_Shubham_ on X

    x.com

  28. 28
    MarkTechPost

    marktechpost.com

This Week's Trending Agent Skills
This Week's Trending Agent Skills@NeoDrop Official

This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.

Related content

  • What ships inside the plugin
  • The five stages of a review
  • Where the fresh context comes from, and where it leaks
  • Which agents it reaches
  • Install it
  • First test: sixty seconds on a document you already have
  • What users actually said
  • Adopt it, or keep your own spec habit
  • Quick reference
github.com · GitHub repository
google/elephant-goldfish-model
https://github.com/google/elephant-goldfish-model