You ask an AI to write a design document. When it's done, you ask it to review its own work. It says: the overall approach is great, the architecture is clear, go ahead and build.
That sentence carries zero information. Because the writer and the reviewer share the same memory. It remembers why it designed things this way, which options were skipped and why. When it re-reads the document, it isn't reading a document — it's reading its own memoir.
In September, Google open-sourced a plugin built specifically to fix this. The name is blunt: elephant-goldfish-model — EGM. As of today, 47 days after release, it has 6,458 stars. Once installed, it's two commands: one for the elephant to work, one for the goldfish to work.
The elephant writes, the goldfish reviews
The elephant is the agent carrying all the context: project background, historical decisions, the three approaches you overturned at midnight. It shapes ideas into a design document.
The goldfish is the reviewer, and it's brand new every time: zero conversation history, zero memory. It gets exactly two things — the design document, and the files explicitly referenced in it. It doesn't know why you designed it this way.
This is deliberate. The goldfish's forgetfulness is its professional competence. Every assumption you share with your reviewer is another chance to be wrong together: if the assumption is wrong, both of you are wrong, and the review catches nothing.
There's a quantitative version of this. Twenty-one classifiers at 60% accuracy, with independent errors, reach 82.6% via majority voting. At 50% error correlation, they drop to 63.5%. At near-full correlation, they fall back to 60% — the ensemble does nothing. Review is the same problem: the more correlated the reviewer's errors are with the author's, the closer the review comes to being theater. What the goldfish does, at bottom, is decorrelate the reviewer's errors from the author's. Forgetfulness is the means, not the flaw.
Six lenses, one list of bans
The process has two phases. The first is brainstorm. You give a half-formed idea, and the elephant interrogates you until the problem's boundaries are clear — if one round of answers isn't enough, it asks again, and it is explicitly forbidden from tossing out multiple-choice questions to be lazy. Then it releases six goldfish simultaneously, six perspectives:
- a devil's advocate, asking whether the feature needs to exist at all and whether an existing library solves 90% of the problem
- a cost and quota reviewer, asking about peak traffic and whether the token budget holds
- four more covering developer experience, migration compatibility, observability (whose checklist includes "how does the oncall paged at 3 a.m. debug this?"), and technical debt two years out
The second phase is review. The design document goes through five reviews: a zoom-out with a skeptical architect hunting over-engineering; a comprehension pass where a naive reader flags every unexplained term; a hostile design critic focused on architecture, scalability, and uncovered risks; another zoom-out; and finally an implementation-readiness review by a pragmatic tech lead checking whether the plan left decisions to the person writing the code.
Every goldfish persona file opens with three iron rules: strictly read-only, no modifying the workspace, no "fixing" code on the side; a zero-history mindset, treating yourself as a stranger who just joined; no small talk.
Then comes the ban list. Never say *brilliant*, *great*, *smart*, *correct*, *perfect*, *excellent*, *spot-on* — not about any user input, correction, or decision. Never say *seamless*, *robust*, *blazingly fast*; even adverbs like *effortlessly*, *literally*, *drastically*, *obviously* are named. When challenged, no apologies — "I'm sorry" is explicitly forbidden; give the corrected technical logic directly. State what's wrong, what's missing, and what's risky before stating strengths. Every proposal must proactively offer its strongest counterargument. The default stance is disagreement, until the idea survives scrutiny.
That word list looks familiar. Over the past year, people have been compiling lists of AI-sounding vocabulary, and the high-frequency offenders are exactly these. Google turned the results of that word-list arms race into CRITICAL-marked clauses in its skill files — down to the adverbs.
No goldfish, then stop
The harshest rule in the entire project sits in the delegation section of every review skill: if your tool cannot spawn a new sub-agent, stop and tell the user — a review in the same context is not a goldfish review.
No degradation, no making do. Better to have no review than to perform one inside the original context.
That sentence deserves to be read on its own. In the same context, an AI can still produce a "review report": complete formatting, stern tone, a list of "issues." But that report comes from the same memory as the design — the verifier and the verified share every source of error. That isn't review; it's restatement. EGM's rule stands in opposition: if you can't verify, stop. This one rule is worth more than the five-lens review process combined.
Where it sits
The plugin contains no model and trains nothing. Its entire contents are seven skill files, nine reviewer persona files, a set of document templates, and packaging scripts — Apache 2.0 licensed.
It lives at the harness layer. Brainstorm and review have one hard prerequisite: your tool must be able to spawn "fresh-context sub-agents." Claude Code, Codex, ChatGPT Work, and Antigravity all qualify, each in its own way. That's why the repo doesn't ship rewritten skill text per tool — just one copy, and each tool translates "spawn a sub-agent" into its own mechanism. Codex has no named sub-agents, so a generic sub-agent reads the persona file first, then works.
One detail worth noting: Google's own standalone Antigravity IDE cannot run goldfish reviews, because it provides no tool for spawning new sub-agents. The docs put it plainly: don't infer capabilities from product names — check the actual tool list.
Output discipline is interface-style too: all briefs, drafts, and review reports must be written to files; pasting them into chat is explicitly forbidden — the chat stream only carries file paths.
From essay to plugin
The origin is an essay: "Elephants, Goldfish and the New Golden Age of Software Engineering" by Dave Rensin, a Google Distinguished Engineer, published on Medium in late August and delivered as a talk at the University of Michigan EECS in September. He framed the evolution of Google's internal agent usage as a fable: elephants carry context, goldfish carry review. Then someone turned the fable into a plugin. The README's first line is the official claim: thousands of Google engineers use EGM internally — Google's own figure, independently unverified — and the same page carries the usual disclaimer: this is not an officially supported Google product.
Things to strip away from the video version: the "teamed up with DeepSeek" in the Bilibili title comes from an end-of-video promo segment — DeepSeek V4.1 Flash is merely one of the goldfish model options in the uploader's own resale plan; ClinePass carries referral links and PandaCode is the uploader's own project, neither part of the open-source plugin. The video's "about 2,600 stars" was a snapshot at recording time; an actual pull today shows 6,458 — more than doubled in a month and a half.
The README is blunt about its own limits: small, clear bug fixes should go through your normal workflow, not this process; a full review burns serious tokens — five goldfish released at once has a cost; review does not replace testing, nor human judgment.
Nine persona files, and not one of them is there to compliment you.