An experiment in spec-driven, clean-room implementation: two coding agents, given the exact same SPEC.md, independently building the same two small apps with no visibility into each other's work.
This is exploratory / not a production project — treat it as a reference for the eval, not a template to build on top of.
Two apps, each fully specified in a single SPEC.md:
- poker — a real-time planning-poker room (SSE, Redis, optimistic concurrency)
- retro — a real-time retro board (same constraints, different domain)
Two agents each implemented both apps from the identical spec, with no access to the other's output:
| Directory | Built by |
|---|---|
gemini-agy/ |
Antigravity CLI (Gemini-based) |
grok/ |
Grok Build (xAI) |
Each app directory is self-contained (own package.json, SPEC.md, tests/, public/).
The SPEC.md files here are copies of the specs published in eswat2/proto-specs (live site) — poker/SPEC.md and retro/SPEC.md. That repo is the canonical source: it explains the goal of publishing specs instead of code, and has a ready-to-use cleanroom-build prompt for anyone who wants to reproduce this experiment with a different harness or model. This repo (spec-cleanroom) is one such reproduction, comparing Antigravity/Gemini and Grok Build/xAI; proto-specs' own README documents a separate Claude Sonnet 5 and Grok 4.5 run.
cd gemini-agy/poker # or grok/poker, gemini-agy/retro, grok/retro
npm install
npm start # starts the server (auto-starts/expects a local Redis — see the app's own README/SPEC)
npm test # runs that app's test suite against RedisEach app requires a local Redis instance (redis-server) unless it documents auto-bootstrap (see gemini-agy/poker/start.sh).
docs/AGY_VS_GROK_REVIEW.md— a detailed comparison of the two implementations (architecture, correctness, hardening, test depth), originally drafted by Grok and independently re-verified against source by Claude as a neutral third party. Includes a verification appendix logging what was confirmed vs. corrected.docs/SPEC_AS_FORCE_MULTIPLIER.md— a reflection on how much of the implementation speed/quality is attributable to having a fully implementation-derivable spec, with a caveat section on the limits of that claim.
MIT — see LICENSE.