You know the routine. You give an AI one sentence, it builds something that looks almost right, and then you spend five, six, seven rounds fixing it — each fix breaking something else. The real problems only show up after real people have already seen the work.
Google's new Elephant-Goldfish Model (EGM) attacks exactly that. It's a free, open-source plugin for AI coding tools like Claude Code, Codex and Antigravity, and its whole trick is embarrassingly simple: before the AI builds, a second set of AI eyes — with no memory of you or your project — reads the plan the way a stranger would.
Elephants remember everything. That's the problem.
The elephant is your regular AI assistant. It knows your project, your style, your past decisions — so it writes beautifully detailed plans. But all that memory hides the gaps: the elephant silently fills in things you never actually said, so the plan looks flawless to both of you.
A goldfish is the opposite: a brand-new AI with zero history, handed only the plan. It doesn't know what you meant. It only knows what the plan says. And it finds the holes in about 40 seconds.
The loop is three steps: the elephant plans, the goldfish review, you decide which fixes matter. Then the AI builds — once.
Why does this work so well? Because your real audience is a goldfish. The Instagram scroller has never heard of you. The person seeing your ad gives it one second. The first-time visitor to your website has no idea what you sell. EGM lets you test your work on a simulated stranger before real strangers see it.
The 19-line plan that "looked fine"
The video's best example is a short plan for a clinic: an AI agent that replies to new patient enquiries, sends a text, answers questions and books appointments. Nineteen lines. Looked perfect.
Four goldfish reviewers — one checking simplicity, one reading as a total beginner, one attacking the design, one asking "could anyone actually build this?" — all returned the same verdict: not ready. Here's what they found:
- No emergency rule. If a patient messaged "I'm in pain," the AI had no instructed response.
- Double-booking. Two people could grab the same appointment slot at the same time.
- No test run, no off switch. The plan switched the agent on for every customer on day one.
- And one reviewer alone spotted two nastier traps: spammers flooding the contact form, and visitors typing instructions into the message box to manipulate the AI into saying things the clinic never approved.
Those are launch-day disasters — caught in about 40 seconds, before a single line was built.
Five builds, same prompt, one difference
Each build below was made twice with the same AI and the same prompt. The only change: goldfish reviewed the plan first.
- 3D website for a coffee subscription. The reviewers caught that the sign-up buttons did nothing, that visitors whose phones can't show 3D would see a blank space, and that the plan never specified the prices or FAQs — so the AI would have invented them.
- Instagram carousel about AI tools. The final slide broke the plan's own rule (it told readers to "pick one" without saying one of what), some text would have shrunk to 9 pixels on a phone, and a guess was stated as a fact. Fixed, the carousel became something people actually save and share.
- Nine ad creatives for a skincare serum. An extra "cold scroller" goldfish — one that had never heard of the brand — scored the direct ads 2.0 out of 10: the same ad nine times, with the offer buried or missing in eight of them. After review: 5.7 out of 10, with the offer readable in half a second.
- A 15-second promo video for a fitness app. Reviewing the storyboard caught an opening claim so big it could get the ad rejected, a calendar viewers could read backwards, and a coaching moment with no cause and effect. All fixed before a single frame existed.
- A lead-generation agent. The lead score was quietly rigged — 30 of 100 points handed to every business no matter what. People who said "no thanks" weren't tracked, so they'd be contacted again the next day. And the AI could invent facts in its outreach messages. Three catches, three fixes.
Same AI. Same prompts. The only difference was a fresh set of eyes reading the plan first.
How to try it (two commands)
EGM is free and open source. In Claude Code:
claude plugin marketplace add google/elephant-goldfish-model claude plugin install egm@egm
Then the workflow is /egm brainstorm (talk the idea through), /egm draft (the elephant writes the plan), /egm review (the goldfish attack it), pick the fixes that matter, and build. It also ships extra skills — resolve walks you through findings one by one, review loop builds a custom reviewer panel, and drift report compares your plan against what actually got built.
Three doubts, answered
- "I'm not a coder." The plans are written in plain words, and everything in the tests — website, carousel, ads, video, agent — was made by describing it in plain words too.
- "Won't this slow me down?" For a typo or a headline tweak, skip it. For anything that would normally take five rounds of back-and-forth, 40 seconds of review replaces all of them.
- "Does it fit my business?" If strangers look at what you make — a page, a post, an ad, an email — then yes. Every customer starts as a goldfish.
Try it this week
Install EGM. Pick one thing you're about to make anyway. Run brainstorm, draft, review — and fix the three findings that matter most. Then build it and compare it with your old way, side by side.
The elephant still does the building. The goldfish just make sure it builds the right thing.
Based on "Google's FREE Elephant Goldfish AI Agents!" by Julian Goldie SEO — watch the full video: youtu.be/fo9fyl-xJQM · EGM (free, open source): github.com/google/elephant-goldfish-model · egm.dev