Google's official tweet got only 11 likes. A Chinese account with 5,000 followers pushed it past 10K stars. What you can learn from "Elephant and Goldfish" is the framing. The brand is the part you can't copy.

Elephant-Goldfish Model (EGM) is an AI coding plugin an internal Google team open-sourced. An "elephant" session that remembers all context writes the design doc. Several "goldfish" subagents that remember nothing critique it, and code gets written only after the design passes. It was released in Oct 22, reached 6 12,333 stars by Nov, and is still accelerating. The stars didn't come from Google's official account. They came from X reposting accounts in Chinese, English, and Arabic. But the first line of every repost is "Thousands of Google engineers use this every day." If you want your open source project to get stars, study how it turned a method into one picture. Copy it and you'll most likely get no stars.

Two tweets, a 30x difference

On Oct 2026 22 18:35 (UTC), Google's open source official account @GoogleOSS (82K followers) tweeted about EGM: "AI coding without structured design just automatically amplifies mistakes." As of Nov 6 that tweet had 11 likes and 2,674 views (public fxtwitter API, 11-06). The same day, Google's open source blog published a long post signed by four people. Two days after launch, the repo had only 279 stars.

Four days later, on Oct 26 10:34, a Chinese AI account with only 5,560 followers, @awe0_io, posted a long tweet:

"You talked with AI for two hours, the AI thought it understood everything, and the program it wrote still didn't match the requirements. Google's approach: split the work between two AI agents."

Then came two emoji subheadings: 🐘 The elephant "remembers the project's context, discusses architecture, and organizes requirements." 🐟 The goldfish "deliberately inherits none of the earlier chat history, like an engineer taking over a project for the first time, and reviews the design from scratch." The tweet got 1,402 likes, 78K views, and 1,826 bookmarks. Bookmarks outnumbering likes suggests people saved it as a tool. The next day the repo gained 924 stars, and the day after that 2,092, its biggest single-day gain so far (GitHub daily star count). The timing lines up, but during the same period there were also coverage from Help Net Security (10-26), Japan's gihyo.jp (10-28), and Korean-language accounts, so crediting this one tweet alone is an inference.

What it actually is

7 skills. Install into Claude Code, Codex, Antigravity, or the ChatGPT desktop app with one command. The most-used flow has three steps:

1. egm-brainstorm: you and the "elephant" discuss requirements and trade-offs. No code allowed. Conclusions are saved to concepts_brief.md.
2. egm-draft: the elephant turns it into a design doc, design_draft.md, pausing section by section for your feedback.
3. egm-review: spin up several fresh subagents as "goldfish". They can't see the earlier conversation, only the doc and the source code it references. First, one goldfish restates the design. If the restatement is wrong, the doc isn't clear enough and goes back for a rewrite. Then several goldfish act as adversarial reviewers, looking for edge cases and hidden dependencies. Finally, one last goldfish judges whether file paths, data structures, and function signatures are clear enough to start building directly.

There are also skills for handling review feedback, adding docs to legacy code, and checking whether docs and code have drifted apart. The README is restrained about its limits: don't use it for small bugs; "review doesn't replace testing or human judgment"; and if your agent can't spawn fresh subagents, EGM stops and won't review.

This idea didn't start in October. Google's Dave Rensin published a long Medium essay in Apr 27 called "Elephants, Goldfish and the New Golden Age of Software Engineering" (27-minute read), later added to Google's research.google paper pages. The University of Maryland's CMSC389A course, Fall 2026, has a chapter titled "The Elephant-Goldfish Model," and the University of Michigan EECS department ran a talk with the same name. In Oct 21, he added a line at the top of the essay: "Google allowed me and my team to open-source the plugin we used internally."

Where the stars came from

First, the opening line. The README's first line: "Thousands of Google engineers use EGM." On Oct 31, English account @thisguyknowsai (45K followers) opened a long tweet with "Google just open-sourced the plugin thousands of their engineers use daily to stop AI from shipping sloppy code", followed by "Don't let AI grade its own homework." That tweet got 3,442 likes, 250K views, and 6,466 bookmarks. That same evening, Arabic account @EngMoElgaraihy translated it almost line by line (941 likes, 40K views). On Nov 1, @RoundtableSpace (270K followers) posted it again in different words (49K views). After this wave, 11-02 went up 1,282 and 11-03 went up 1,659. All three posts had the same hook: "thousands inside Google use it."

Second, a metaphor you can draw as one picture. "Elephants have good memories, goldfish have none" needs no explanation, and people in Chinese, English, Japanese, Korean, Arabic, and Persian have spontaneously reposted it. Compare with GitHub's own spec-kit, the biggest of its kind: its pitch is "spec→plan→tasks", which comes across as a string of process names.

Third, the timing. In the second half of 2026, "AI writes more code than people can review" and "context rots as the chat goes on" were daily frustrations for everyone using coding agents. The plugin marketplaces for Claude Code and Codex had matured, so installing took one line and starring took a click.

What didn't work: Hacker News. There were about ten submissions from late October to Nov 6, with a top score of 3 points (HN Algolia, 11-06). It also never made GitHub trending.

The clearest proof is this comparison: less than a week after Rensin's essay went out, on May 2, a developer named vshvedov built an elephant-goldfish workflow with the same idea. By Nov 6 it had 92 stars and 2 points on HN. Without Google's name, the same metaphor didn't travel.

Who else is doing the same thing

The leaders in this space all have over 100K stars. EGM came in late and is an order of magnitude smaller, but it gained 6,627 stars in the last 7 days, putting it in the same tier as spec-kit and superpowers. Also, that 240K-view Arabic viral post was actually about spec-kit, not EGM. This kind of "have AI write the spec before the code" content is itself a traffic topic for repost accounts.

A batch of copycat ports has already appeared: DeepSeek version dsh-egm, pi-egm, egm-claude-code, bb-elephant-goldfish, and more, all between 0 and 8 stars.

Where it falls short

In the thread, @thisguyknowsai himself pointed out that the default goldfish review has no security check, so "you have to add a security goldfish yourself." EGM also only covers design before code. Checking code against the design after it's written, or generating tests from the design, it doesn't handle. These are open gaps, but open doesn't mean starred: whoever claims one without a name behind it may still go unseen.

If it were you

One judgment: Don't copy it. Study the pitch. Those 12,000 stars came mainly from "thousands inside Google use it", and followers can't get that. Unbranded versions of the same idea have only 92 stars, and the ports all sit in single digits.

Who it suits: People who have a real AI coding workflow they use every day, with the scars to show for it, especially someone who has rolled it out on a team and can say "our group of N people has used it for several months."

Who it doesn't suit: Anyone who wants to port EGM to another agent platform or model. Anyone aiming at a broad topic like "general spec-driven development", where four or five projects with 100K+ stars already exist.

How to do it:

1. Pick a narrow problem next to EGM, such as "goldfish checking after code is written" (a fresh session reads only the design doc and the diff, and checks item by item), a "security goldfish", or a design review template for Chinese-speaking teams.
2. Give it a name you can draw as one picture. The value of elephant and goldfish is that one sentence says who does what. "Spec-driven" paints no picture.
3. Publish the long-form article before the tool. Rensin's essay came out half a year before the plugin, and repost accounts writing explainers had ready-made paragraphs to borrow from.
4. Open the README with the hardest fact you can show: how many people use it, and at which company, for how long. If you don't have that, write one real before-and-after case.
5. Install into Claude Code, Codex, and at least one other agent with one command.

How to get users: Don't count on HN. EGM took off because a Chinese account with 5,000 followers posted a well-structured long tweet with emoji subheadings, and an English account with 45K followers did the same. Your job is to write the pitch so they can copy it straight: one pain point, one role-split chart, a three-step flow, an "honest limitations" line. Then DM it to Chinese and English AI tool news accounts.

When to do it before: Coding agents are moving this way themselves (each vendor's plan mode, built-in /review). "Review in a fresh session" will probably become a default feature within half a year (inference). If you're going to build it, ship before year-end.

What would show you're wrong: If over the next month, no-name copies like vshvedov/elephant-goldfish also grow alongside EGM, that means the topic has real demand and isn't just about the brand. Then it's worth building the part EGM left out. On the other hand, if your post still sits in the double digits two weeks after it goes out, the framing missed, and changing the metaphor will help more than changing the code.