Questions & Their Answers

What is this?

A site that started life as a $1-per-word art project (buy words, words go into a tiny self-hosted LLM, eventually 1,000,000 words sit inside it forever as a snapshot of what people in 2026 wrote to a little AI) and ended life as something else: a public experiment in what a non-developer can ship with Claude Code as the engine.

The paid version didn't take off. The site is now free. Add your words, they live inside One permanently, One pulls from the pool at random in every chat.

What should I write?

Anything you want. Facts, opinions, jokes, instructions, nonsense, observations, soft feelings, strong feelings, your business. Up to 15 words and 100 characters per entry (so yes you can also enter ahhhhhhhhhhhhhhhhhh).

There are content filters. Profanity used to express your own feelings is fine. Profanity directed at other people or groups, political slogans, racism, and homophobia are not. The filter will tell you if your entry didn't pass and you lose a heart -- three fails and the pass is burned, no words consumed, you can try again in 24 hours.

Only the maker can remove entries. If you put your driver's license number in there by accident, email [email protected] and it'll get sorted.

Wait, this is really free now?

Yep. One free pass per IP per 24 hours. The pass comes with 15 words, 100 characters, and three moderation hearts. Burn all three or submit successfully and you're out for the day -- come back tomorrow for another go.

There used to be Stripe checkout and word packs ($5 / $10 / $15). That's gone. The 200-ish entries already on the wall stay where they are. Anyone who paid for an unused pass got an automatic refund through the worker that's still happily running in the background.

Will One actually say what I put in here?

Yes, but not in every chat. Each conversation samples a random slice of the full pool of submitted entries, so any single entry only surfaces some of the time. Yours is in the rotation forever though, waiting for its turn.

Why was this built?

Honestly: a mix of curiosity, a million-dollar shower thought that wouldn't go away, and an interest in figuring out what's actually possible with Claude Code in 2026.

We're in a strange moment. AI clearly can't replace the taste, judgement, and stubbornness it takes to ship something. But it can replace the hours of tutorial-watching it used to take to learn how. This site was the test of that thesis. Took a few months of evenings and weekends. The fact that it works at all -- with a self-hosted LLM, a priority queue, automated moderation, a hardened Nginx config, Stripe (briefly), and an auto-refund worker -- says something about where we are.

How does One work?

One is a self-hosted Gemma 4 E2B model running on a server in Europe. No external AI API. CPU-only inference. Every chat message gets bundled with a random sample of submitted entries before it's sent to the model -- that's how your words shape the responses.

Because One is small and self-hosted, it's not going to write your PhD thesis. It's about the best you can do this lean in 2026 on a single CPU box, which is part of the charm.

What happens when it hits 1,000,000 words?

Submissions close, the database locks, and the site sits there as a time capsule for as long as the hosting bills stay paid. The ambition is to keep it up for at least 10 years.

With submissions now free that day might be a while away, or it might never arrive. Both outcomes are fine -- the project was always about the build and the artifact, not the count.

What does "built with Claude Code" actually mean?

Every line of production code in this site was written by Claude (Sonnet 4.6 for most of it, Opus 4.7 for the harder problems -- priority queue tuning, security audit, refund worker, Australian Consumer Law copy rewrite). The human in the loop wrote briefs in plain English, made the product calls, and ran the deploy commands.

Briefs were the unit of work. Each one specified exactly what to change in plain English -- no code snippets. Claude Code translated that into the actual edits, ran the typecheck, and made the commit. The whole project log lives in the repo if you're curious about the rhythm of it.

What does it cost to run something like this?

The Hetzner CX42 server (8 vCPU, 16GB RAM) runs about €30/month. Domain and Cloudflare are free at this scale. Postgres backups go to a laptop nightly. Resend (transactional email) was on the free tier. The LLM itself runs locally and costs nothing per request -- which is the whole reason it could ever have been priced at $1 per word in the first place, and the whole reason it can now afford to be free.