whyc/whyc Public

Code Issues 12 Pull requests 3 Actions Projects Wiki Security Insights
sejun-kim convergence reached: spec-fit 96% on synth-jd-08 main a3f91c2 · 14 minutes ago · 147 commits
.github/add issue templates for spec-fit failures2 days ago
agents/extractor + designer + judge ADK graph14 minutes ago
phoenix-mcp/trace queries + LLM-as-judge prompt3 hours ago
preview-runtime/Next.js template + design-token contractyesterday
LICENSEApache-2.0 (OSI-approved)36 days ago
README.mdwhile YC hires, we ship14 minutes ago
cloudbuild.yamlCloud Run deploy + Secret Manager wiring2 days ago
README.md

WhyC — while a certain orange accelerator hires, we ship.

licenseApache-2.0 OSIapproved stackGemini ADK obsPhoenix MCP deployCloud Run spec-fit96% trackArize buildpassing

You give us a job-posting URL from workatastartup.com. We give you a deployed Next.js preview on Cloud Run inside 24 hours, then a self-improvement loop keeps regenerating the weak flows until the deployed app actually matches its own spec.

Their moat is a closed batch. Ours is git clone. Every layer below is open source — including the judge that grades us.

Quick start

# 1. clone — yes, the whole satire is forkable
git clone https://github.com/whyc/whyc.git && cd whyc

# 2. point it at any YC job post (or paste a company description)
export GOOGLE_API_KEY=...   # Gemini ADK
export PHOENIX_ENDPOINT=... # Arize Phoenix MCP
npx whyc ship --from https://workatastartup.com/jobs/<id>

→ deployed:  https://preview-7f3a.run.app
→ spec-fit:  71%  (loop running in background...)

How it works

1. Extract Gemini → Spec 2. Generate tokens + Next.js 3. Deploy Cloud Run URL 4. Judge Phoenix MCP + LLM 5. Regen weak flows only until spec-fit converges

Why this is open source (and why it has to be)

  • The hackathon Stage-1 gate requires an OSI-approved license. Apache-2.0, no negotiation.
  • Phoenix is open source. Gemini ADK examples are open source. Next.js is open source. The only thing left for an accelerator to gate-keep is a closed batch — and that is the joke.
  • Every regenerate decision is logged through Phoenix and reproducible from this repo. The judge prompts live in phoenix-mcp/judge/*.txt. Disagree with our scoring? Open a PR.

Convergence, in public

Latest run on a synthetic JD (no real company referenced):

iter 1 · 71% iter 2 · 84% iter 3 · 92% iter 4 · 96% ✓

Full traces: phoenix-mcp/runs/synth-jd-08.jsonl · LLM-judge diff: agents/judge/diff-iter3.md

While they hire, we ship. A certain accelerator's job board has a posting from a Demo Day company that is now 197 days old. We took a synthetic version of that JD and shipped a working preview in 23h 41m. The repo you are reading is the receipt.

Contributing

We accept PRs that (a) add a new judge rubric, (b) tighten the design-token contract, or (c) port the runtime template to a non-Next.js framework. See CONTRIBUTING.md. Be nice in issues. Be ruthless in code review.

License

Apache-2.0. © 2026 WhyC contributors. Not affiliated with, endorsed by, or naming any specific accelerator. Categorical satire only.

Open #42 · self-improvement loop didn't catch the missing /pricing flow on iter 2 opened 6 hours ago by indie-hacker-99

Repro: ran npx whyc ship --from synth-jd-04. Spec extracted a /pricing flow but the deployed preview shipped without it. Phoenix trace shows the judge marked it 89% (above our 85% gate) on iter 2 and exited the loop. Suggest tightening the gate to 92% when spec.flows.length > 3.

Labels: judgegood first issue