EGM: Google's New Way To Use Claude Code Is Insane

15:543 min readAI LABS10/25/2026Computing/Software

Key Takeaway

Google's Elephant Goldfish Model prevents faulty AI code generation by enforcing rigorous spec-driven planning and blind multi-agent critique before any implementation begins.

Highlights

  • The Elephant Goldfish Model is a spec-driven development framework designed by Google engineers to plan and review software features before an AI agent writes code.

  • The framework separates agents into two core types: elephants that maintain project history and context, and goldfish sub-agents with zero context that aggressively critique plans and catch hidden flaws.

  • Running the bootstrap command on an existing application maps every folder containing code and generates individual readme files documenting folder rules and behavior.

  • The review phase runs across five distinct rounds using sub-agents to analyze edge cases, scalability issues, and architectural alignment without being influenced by prior chat history.

  • A drift report command compares the codebase against established readme documentation to identify code discrepancies and actual bugs after features are implemented.

Timeline

  1. Core Philosophy of the Elephant Goldfish Model

    Traditional coding agents build features immediately and expose problems only after implementation. This model splits responsibilities between context-aware elephant agents and isolated goldfish sub-agents. Goldfish agents examine plans to find simpler implementations, check clarity for unassisted coding, and evaluate scaling limits under heavy user traffic.

    • The framework uses spec-driven development to map out a complete plan before an agent builds an application feature.
    • Elephant agents maintain full project history and collaborate with users to brainstorm and write initial design documents.
    • Goldfish sub-agents operate with zero chat context to critically evaluate proposals from multiple distinct perspectives without bias.

  2. Installation and Project Setup via Bootstrap

    Users install the tool using repository commands or direct prompt instructions. Bootstrapping an application allows sub-agents to document code architecture across every folder. For instance, testing on a demo application generated 28 new readme files and updated the central project documentation to establish ground rules for future modifications.

    • The model is free to use from Google's GitHub repository and supports major AI coding tools including Claude Code and Codex.
    • Installing the repository creates designated agent folders, skill scripts, and framework documentation.
    • The EGM bootstrap command systematically inspects code folders and generates individual readme files outlining internal rules and responsibilities.

  3. Brainstorming and Concept Brief Generation

    Adding a new feature begins with the brainstorm command. The system interrogates the user to resolve ambiguities, critiques any existing drafts, and passes the refined concept to isolated sub-agents. These sub-agents grade the proposal against written rules and compile a consolidated concept brief saved in the main project directory.

    • The brainstorm workflow evaluates existing plans, asks targeted questions, and structures a clear approach for new features.
    • A goldfish review process initiates multiple background sub-agents that analyze the initial proposal independently.
    • The resulting concept brief outlines the feature purpose, potential roadblocks, and explicit execution steps.

  4. Designing, Reviewing, and Managing Code Drift

    Once the concept brief is approved, drafting generates technical implementation plans and alternative designs with documented tradeoffs. Sub-agents then review the plan across multiple sequential rounds to catch oversights. After implementation, the drift report command cross-references the live repository against readme files to catch unintended code modifications and logic errors.

    • The draft command creates detailed design specifications, critical user journeys, and technical alternatives considered.
    • A five-stage review phase subjects the technical plan to rigorous multi-angle scrutiny before changing its status to proposed.
    • Drift reports analyze the active codebase against project documentation to flag unexpected deviations and logic bugs.

Description

This Claude Code workflow from Google tears apart your plan before your agent builds anything. We show how to use Claude Code agents with zero context to review it, the Claude Code setup to install it, and how it caught a real bug in our app. If you build with AI coding tools, watch this first. Create your next app promo with Hedra at https://hedra.link/ailabs and use code AILABS50 for 50% off your first month. Elephant-Goldfish Model on GitHub: https://github.com/google/elephant-goldfish-model Community with All Resources: https://ailabspro.io/?v=xxf6D9jynFw The Roundup, our daily newsletter covering the AI stories that matter. Join now: https://www.theroundup.so/ Got a software idea but don't have the team to build it? Autometa turns your ideas into production-ready software. Check out https://autometa.dev/yt to see what we can build for you. Usually when you ask your agent to build something, it just builds it, and you only find the problems afterwards. The Elephant-Goldfish Model is a Claude Code workflow built by engineers at Google that turns that around. It's based on Dave Rensin's article "Elephants, Goldfish, and the New Golden Age of Software Engineering," and it's free on Google's GitHub. It's a Claude Code agentic workflow with two roles. The Elephant remembers everything about your project. It talks through your idea with you and writes the detailed plan. The Goldfish are Claude Code agents with zero context. They only see the plan and your app's code, never your chat. That's why clearing the context matters in this Claude Code workflow: an agent that saw your conversation tends to agree with itself, and a Goldfish can't. You're the third role. The review gives you lots of notes on purpose, and you make the final calls. In this Claude Code agent workflow tutorial, we run it on a demo fashion app and add a page that shows sellers which sizes are selling and when to restock: - Bootstrap: maps your app one folder at a time and writes a README for each one (28 in our case), plus the fix so Claude Code actually reads them - Brainstorm: asks you questions until your idea is clear, then six Goldfish review it and write a concept brief - Draft: writes the design plan one part at a time, including user journeys, the technical plan and the alternatives it considered - Review: five rounds of Goldfish review before anything gets built - Drift: checks your app against the plan later on. It caught a real bug in ours: shop owners could delete products that had already sold Claude Code workflow vs Codex? You don't have to pick. The repo has an install command for Claude Code, Codex and other agents. We also cover one line the Claude Code setup wrote on its own that you'll want to check, because it says subagents can't start their own subagents, and the new version of Claude Code can. If you've wondered how to create a workflow with Claude Code that plans before it builds, this is one way to do it. It replaces the usual Claude Code creative workflow of "ask, build, fix later" with a plan that gets torn apart first. The Claude Code agents workflow automates the review with subagents and skills, so it's a practical example of AI automation and AI agents working together on your own app. Is it the best Claude Code workflow? For large features where a mistake would cost you a lot, it went deeper into edge cases than the other approaches we've tried. For a small fix your agent can do in one go, Google says it isn't worth it. Whether you're new to Claude AI or already using Claude Code skills, this video shows you how to use Claude Code to plan properly. 00:00 Intro 00:47 What It Is 01:43 Elephant vs Goldfish 03:02 Why Not Plan Mode 04:54 Install 06:07 Fixing CLAUDE.md 07:08 Bootstrap 09:29 Sponsor 10:18 Brainstorm 11:32 Draft 12:42 Review 13:50 Drift Report 14:35 When to Use It Hashtags: #ai #claude #claudecode #aiautomation #claudeai #aiagents #howtouseclaudecode #claudecodeskills

Write about this video