google/elephant-goldfish-model

What it solves

Writing code from weak or untested specifications leads to rework and drift. EGM tackles this by separating the drafting and the stress-testing of design specs: deep-context agents explore the problem and write a thorough specification, then stateless subagents attack that spec from many angles before any code is written.

How it works

EGM defines two AI agent roles and a human role. The Elephant agent loads the full system context, architecture, and user journeys, and runs structured brainstorming and drafting commands. The Goldfish agents are isolated, zero-context subagents that read only the draft spec and live source code, then challenge it with specific personas (Design Critic, Naive Reader, Tech Lead, etc.). The human owns the final architecture decisions. Workflows start with /egm-brainstorm to idea and stress-test concepts, then /egm-draft turns the concept into a design doc under docs/designs/, then /egm-review runs a multi-stage reviewer swarm (Zoom Out -> Comprehension -> Critic -> Readiness), with additional commands for resolving findings, detecting drift between specs and code, and bootstrapping documentation.

Who it’s for

Engineering teams and individual developers who use AI coding agents (like Claude Code, Codex, or Antigravity) and want to produce verified design specifications before implementing features. It is especially useful for projects where requirements are complex or where design-review discipline is needed.

Highlights

• Open-source plugin for Spec-Driven Development with AI agents (Apache 2.0)
• Deep-context Elephant agents draft specifications using history, architecture, and user journeys
• Zero-context Goldfish subagents stress-test drafts from conflicting personas without conversational bias
• Commands for brainstorming, drafting, multi-stage review, resolving findings, drift detection, and bootstrap documentation
• Human engineers retain ownership of architectural decisions and outcomes