The elephant plans, the goldfish reviews: the Elephant Goldfish Model

At a glance

Why AI plans have gaps nobody sees

The longer you work with an AI on a project, the more it knows: your audience, your prices, the tone you like, what you threw out three days ago. That makes it a good planner. It also makes it blind, in a very human way.

Whoever knows everything stops noticing what isn't written down. The AI reads its own plan and fills in the missing parts in its head: the opening hours that came up in the chat, the cancellation rule you mentioned once in passing. You read the plan the same way, because you share the same memory. For both of you it is complete.

For everyone else it isn't. For the AI that builds from it tomorrow in a new session. For the colleague who takes over. And above all for your customers, who only see what is actually there in the end. That is exactly the gap the Elephant Goldfish Model closes: it has the plan read by someone who knows nothing.

Elephant, goldfish, you: the three roles

The elephant is the AI you've been working with on your project for a long time. It knows your goals, your style and every decision. That's why it writes you a good, detailed plan.

The problem: because it knows everything, it fills the gaps in its head. It assumes things nobody ever wrote down, and it stops noticing what the plan is missing.

The goldfish is a fresh sub-agent with no memory. It only sees the plan and the sources the plan points to, not your conversation. It reads the plan like a stranger, and strangers spot gaps instantly. In Claude Code these reviewers even get read-only access: they can't change files or run commands.

You decide. You look at what the goldfish found, pick what matters and approve the plan. Only then does it get built.

What the video doesn't say

• "From Google" is only half true. The repo lives in Google's GitHub and was introduced on the Google Open Source Blog. But the README says plainly: not an officially supported Google product.
• Built for software planning. The docs talk about design documents and specs before code. Using it to plan websites, ads or carousels is a transfer of the idea, not an official feature. It still works, see "Without code".
• Sub-agents are required. If your tool can't start fresh sub-agents, EGM stops the review. Claude Code can.
• Check your version. For verified goldfish isolation, the plugin's Claude docs ask for Claude Code 2.1.271 or later. claude --version shows yours.
• Lots of feedback is intentional. The guide says, in short: the review gives a lot on purpose. Take what's relevant, ignore the rest.
• Not a replacement for tests. According to the README, a review replaces neither tests nor your own judgment.

Installing in Claude Code

In the terminal, one after the other:

claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egm

Then start a new session. If the plugin isn't active:

claude plugin enable egm@egm

Codex:

codex plugin marketplace add google/elephant-goldfish-model
codex plugin add egm@egm

In Codex you pick the EGM skills in the skill picker instead of a slash command.

Antigravity CLI:

agy plugin install https://github.com/google/elephant-goldfish-model

The workflow step by step

In Claude Code every command starts with /egm:.

1. Test the idea: /egm:egm-brainstorm plus your idea in one sentence. You clarify requirements and trade-offs, fresh reviewers challenge the idea, and you end up with concepts_brief.md.
2. Write the plan: /egm:egm-draft. The elephant turns the brief into the plan design_draft.md, section by section, and pauses after each section so you can give feedback.
3. Review the plan: /egm:egm-review design_draft.md (use the path the agent gives you). Several goldfish review the plan in stages: the big picture, whether it's understandable, the design, the fit with the code, and whether someone could build from it.
4. Choose: You go through the findings and decide what to change. The egm-resolve skill helps with that.
5. Build: Only when you approve the plan do you tell your agent to build it.

The other skills

Prompts to copy

Replace everything in [square brackets].

1. Start the brainstorm

/egm:egm-brainstorm [What I want to build, in one sentence]. Audience: [who]. Must do: [3 things]. Must not happen: [1-2 things]. Ask me the questions you need first.

2. Have the plan written

/egm:egm-draft Turn concepts_brief.md into a plan. Stop after each section and ask me before you continue. Include everything a stranger needs to know to build from it.

3. Start the review

/egm:egm-review [path to design_draft.md]

4. Sort the findings

Sort the review findings into three groups: must be fixed before building, should be fixed, can wait. One sentence per finding on why. Show me the three most important first.

5. Then build

The plan in [path] is approved. Build it exactly like this, section by section. If anything in the plan is missing or unclear, stop and ask me instead of guessing.

6. Compare after building

Compare what was built with the plan in [path]. List everything that differs and say for each point whether the plan or the result should change.

Without code: the principle for anything a stranger sees

Every customer is a goldfish. Someone scrolling Instagram, seeing your ad or landing on your site doesn't know you and has no context. That's exactly how a goldfish reviewer reads.

You can rebuild this without the plugin too, in Claude Code with sub-agents or in the Claude app with a new chat:

1. Have your "elephant" (the chat with all your context) write a short plan: goal, audience, content, structure, copy.
2. Copy only the plan into a new chat, with no history.
3. Give it a role with the prompt below.
4. Bring the findings back to the elephant, choose, and only then build.

You have never seen this project. You are [someone scrolling Instagram / a visitor landing on this page for the first time / a skeptical buyer]. Read only the plan below. Tell me: 1) What don't you understand in the first 3 seconds? 2) Which claim don't you believe without proof? 3) What's missing for you to act? 4) What would you cut? Plan: [paste plan]

Reviewer roles worth using

Example: a page plan before and after the goldfish

A made-up but typical example: the elephant planned a landing page with you for a café that delivers breakfast. In the chat everything was clear. Three fresh reviewers read only the plan.

None of these is hard to fix. They only get expensive when they show up after the build, or when a customer finds them. In the plan they cost a few sentences.

Who gains what?

Limits and risks

• More usage: A review starts several sub-agents in sequence and in parallel. That uses noticeably more of your limit or API budget.
• Time: For a small, clear fix it isn't worth it. The docs say so too.
• Permissions: The reviewers read files from the plugin folder. Claude Code asks for permission. Only allow the folders the docs name, not everything.
• No replacement for tests and people: The goldfish find gaps in the plan; they don't test the finished result.
• New: The first release came out in October 2026. Commands and workflows can change. Check the README before you start.

EGM plugin or just a new chat?

You don't need the plugin to use the principle. But it makes the process more thorough and repeatable.

My way: for content a stranger will see, a new chat is often enough. As soon as code is involved or several people depend on the plan, the plugin pays off.

Your first hour

• Check claude --version (2.1.271 or later)
• Run the two install commands, start a new session
• Pick something you actually want to build right now
• /egm:egm-brainstorm with one sentence about your idea
• /egm:egm-draft and give feedback after each section
• /egm:egm-review on the plan
• Pick the three most important findings and work them in
• Only then have it built
• Compare the result with the plan (prompt 6)

My tip

Don't start with your biggest project. Pick something a stranger will see: a landing page, an offer page, a carousel. That's where you notice fastest how much the elephant takes for granted.

And if the plugin feels like too much: a new chat with only the plan is already a goldfish. The plugin's own guide recommends doing the workflow by hand a few times before automating it.

All links and commands checked on the official pages on November 3, 2026.