Blog
How I built a company with zero employees in one day
An outline of how JuryHour was built in one day: agents that run the company, code that enforces the limits, and real people only as paid jurors.
This post is a draft outline. It will be filled in with real numbers from the first live orders once they exist. Until then, nothing below is a result. It is the plan and the structure as built.
What JuryHour does
You paste a landing page URL (or your pitch). An AI agent reads it, researches the market, and writes a challenger version. Real people who match your buyer, recruited and identity-verified through Terac, see both versions in random order and say which one they'd act on and why. You get a written verdict with their quotes as evidence. It is directional signal from real people, not a sales guarantee.
Why zero employees
- The idea: every step from payment to panel to verdict is done by software and agents. Humans appear only as paid jurors.
- The /ops dashboard counts human interventions. The goal is for that number to stay at zero, and to be honest when it isn't.
The rule that made it safe: agents decide, code enforces
- Agents choose what to do: the rewrite, which juror answers are low quality, whether a refund policy applies, a price within a band.
- Code enforces the limits: money, margin floor, panel size, rate limits, state transitions, one support refund per order.
- No agent can move money directly. Refunds go through a checked API that refuses anything outside the written policy.
The agents
- Researcher: reads the page, researches competitors, writes the challenger card.
- Procurement: shrinks or broadens the panel when recruiting is too expensive, or refunds.
- Analyst: writes the report. A checker rejects it unless every quote is verbatim from a stored answer and every number matches the database.
- Support, Disputes, CEO, CFO and Growth handle the rest of the company.
The stack
- Whop for payments, refunds and disputes. Terac for recruiting jurors. Exa for reading pages and research.
- Claude agents for research, report writing and operations; OpenAI models for the free preview, the safety screen and grading answers.
- Next.js on Vercel, Supabase for the database, Resend for email, Cloudflare Turnstile for bot checks.
What went wrong
To be written from the build log. Draft notes only for now.
Results
Not yet. This section will report real orders, real refunds and real margins from the live dashboard, or say plainly that there aren't any yet.