Under the hood

You do not need any of this to use the experiments.

But if you want to know what is actually running underneath them, here it is, from the shallow end.
  1. PublicWe make things to see what happens.
  2. Curious visitorEvery idea begins as a cheap experiment. Reality decides what happens next.
  3. BuilderAI agents build, deploy, monetize, distribute and measure experiments using reusable skills and tools.
  4. TechnicalClaude · ChatGPT → Skills → Project Nothing → MCPs → internet.

The loop

One number matters more than the others.

Time from “what if?” to evidence. Everything below exists to make that number smaller.
  1. IDEA
  2. BUILD
  3. LIVE
  4. €
  5. PEOPLE
  6. DATA
  7. ?
thenKILLMUTATEDOUBLE DOWN

Show, don’t explain

What one experiment looks like end to end.

The question

Would people pay €3 to turn their Git history into a poster?

  1. exp-0047 created
  2. GitHub repository created
  3. application generated
  4. deployed to Vercel
  5. Stripe €3 checkout created
  6. analytics connected
  7. social campaign generated
  8. posts published — after approval

24 hours later

impressions
1,842
visitors
96
creations
21
checkouts
7
purchases
4
revenue
€12
infrastructure cost
€1.84

Decision

Mutate

Traffic converts. Improve activation and test €5.

This one is invented — a dry run to show the shape of the loop. Every number elsewhere on this site comes from an experiment’s own recorded state, and reads “not measured” when it does not.

Four layers

Agents are disposable. The memory is not.

Project Nothing does not attempt to replace GitHub, Vercel, Stripe or PostHog. It gives agents a coherent way to operate them together.
  1. 01

    Agents

    Intelligence

    Claude, ChatGPT, Codex and whatever comes next are interchangeable operators. No layer below this one may depend on which model showed up.

  2. 02

    Skills

    Behaviour

    Reusable experimental procedures — validate, launch, monetize, distribute, analyze, iterate, kill, graduate. A skill says how to accomplish something, never which vendor endpoint to call.

  3. 03

    MCPs

    Capabilities

    GitHub, Vercel, Cloudflare, Stripe, PostHog, Sentry and the social networks. Existing servers wherever mature ones exist; thin adapters only where necessary.

  4. 04

    Project Nothing

    State & control

    Experiment identity, hypothesis, budget, resources, deployments, distribution, costs, revenue, observations, decisions. Agents are disposable. This is the part that remembers what happened.

Claude · ChatGPT · Agents
          ↓
        Skills
          ↓
    PROJECT NOTHING
          ↓
 MCPs / APIs / Services
          ↓
┌─────────┼──────────┐
BUILD   DISTRIBUTE   MEASURE
          │
        STRIPE
          ↓
       REVENUE

Skills

Procedures, not API wrappers.

A skill describes how to accomplish something. Which service satisfies “deploy it” is a separate question, which is why swapping one out does not require rewriting a procedure.
SkillReachesWhat it does
/validate-ideaREVERSIBLE_WRITETurn a "what if?" into a falsifiable experiment with a question, a hypothesis, a commercial hypothesis, a budget and a timebox — or into a recorded reason not to bother.
/scaffold-experimentREVERSIBLE_WRITETurn a validated idea into a working directory: the contract, the log, the funnel instrumentation and the umbrella badge, with the identity already filled in.
/launch-experimentEXTERNAL_WRITEBuild the smallest sellable version, deploy it, instrument the funnel, and put it in front of a stranger.
/monetize-experimentEXTERNAL_WRITEAttach the smallest real price to the thing and let Stripe carry the checkout. Money is a measurement instrument, not a milestone.
/create-social-campaignEXTERNAL_WRITETurn the experiment into a small narrative with its own distribution hypothesis, drafted for approval — never auto-published.
/announce-experimentEXTERNAL_WRITEPut the approved campaign on a schedule and into the world: media cut from the running product to each network’s format, and every claim checked against the page it points at.
/analyze-experimentREADRead what actually happened — funnel, revenue, errors, engagement — and say what it means, including when it means nothing yet.
/iterate-experimentEXTERNAL_WRITEChange one thing that the evidence pointed at, and put it back into the world.
/kill-experimentEXTERNAL_WRITEEnd it, publish an honest cause of death, and shut down everything it was still paying for.
/graduate-experimentEXTERNAL_WRITEPromote a survivor out of the playground while keeping its origin story permanently public.

Umbrella branding

Experiments do not look like this site.

A portfolio where everything shares a palette reads as one product with a lot of tabs. Each experiment gets to look like whatever it needs to. Four things are shared, and only four.
  1. 01The badge“a Project Nothing experiment”, linking back to the experiment’s page here. It tells a visitor this is a small thing made to find something out, which buys enormous latitude for rough edges.
  2. 02The idexp-0047 visible somewhere, even if only in an HTML comment. It is how a screenshot or a support email traces back to a ledger.
  3. 03The funnel namesimpressions · visitors · activations · shares · checkouts · purchases, exactly, whatever tool is behind them. Bespoke events are welcome alongside; the six are a floor.
  4. 04The honesty line“This is an experiment. It might be gone next month.” Never imply permanence the thing does not have.

No shared palette, typeface, layout, component library or tone. An experiment can be funnier, drier or angrier than the studio, and usually should be.

Autonomy & safety

Permission is a property of the capability.

Every capability an agent can reach declares one of four classes, and one policy engine decides what proceeds unsupervised — rather than each integration inventing its own guard.
  1. READInspect analytics, revenue, logs and engagement. Autonomous.
  2. REVERSIBLE_WRITECreate branches, previews, drafts, instrumentation and experiment records. Autonomous — everything here can be undone.
  3. EXTERNAL_WRITEProduction deployments, social publishing and customer communication. A stranger sees the result, so a human approves it.
  4. SPENDAdvertising and infrastructure purchases. Requires an experiment budget and stays under the approval threshold.

Publishing to a real audience and spending real money both stop at a human by default. An experiment with no budget cannot spend at all — that is not a setting, it is the absence of a number to spend against.

Why

An economically self-sustaining playground for curiosity.

The long-term ambition is not to automate startups. It is to push the cost of trying an idea toward zero, so that experiments continuously produce knowledge and the rare things that work finance everything that comes next.

What accumulates is harder to copy than the code: an audience willing to try whatever gets made next.

See what is runningNothing so far