Elephant-Goldfish Model
Draft specs with AI. Review with fresh agents.
81 followers
Draft specs with AI. Review with fresh agents.
81 followers
EGM helps you draft software specs with AI, then fresh subagents review them without the earlier conversation. Run a drift report to compare specs with code. Open source. Works with Claude, Codex, Antigravity, and ChatGPT.
Free
Launch Team
Render Sandboxes: a computer for every agentGive every agent its own machine. Starts in under a second
Promoted
I’m Oliver. I’ve been working with a small group at Google on Elephant-Goldfish Model. Thousands of Googlers already use it, and it’s now open source!
When you work through a feature with an agent, you build up a lot of shared context. That helps with drafting, but it can also make a review follow the same assumptions.
EGM separates those jobs. Start with an idea and work it into a design with one agent. Fresh reviewers read the design and relevant code without seeing that conversation. You work through the findings, then can run a drift report to compare the design with the code.
EGM has install paths for Claude Code, Claude Cowork, Codex, Antigravity, and ChatGPT. Cowork and ChatGPT have outstanding end-to-end verification gaps. The plugin is free and Apache-2.0 licensed. Your usual model usage costs apply.
Try it on a feature you’re planning and let me know what the reviewers caught, and what they missed.
https://github.com/google/elephant-goldfish-model
Does the drift report work the other way round too? Our standards doc went stale a few months in once people stopped updating it, so a fresh reviewer reading it would have flagged correct code as wrong
The split between drafting with context and reviewing with fresh eyes is the part that stands out to me — that is usually where a review just inherits the same assumptions. Do the fresh reviewers see the relevant code as well as the design, and how do you decide how much code context is enough without pulling the drafting conversation back in?