Since a few of these threads ask what people have actually shipped, here's a full one. https://deduce.studio
I've been building Deduce — a daily deduction puzzle. Cluedo-style logic grid: work out who did it, where, and with what, from a set of clues. New case every day across five difficulty tiers. Free, no signup, no ads, nothing to buy.
It's written almost entirely by Claude, in this pattern: a full spec first (functional, technical, build plan), then three sub-agents run against that spec in parallel — one on the world database and API, one on the frontend and print layouts, one on the puzzle engine. The spec doing the coordinating is the whole trick. Agents working off a chat history instead of a document hand you three incompatible halves.
Stack is deliberately boring: React + Vite, Express, Postgres with raw pg and no ORM. Playwright renders each case to a print-ready PDF for a paperback format. The whole thing runs in a container on a mini-PC in my study behind Cloudflare, so hosting is R0/month.
THE PART AI COULDN'T DO
Generating a puzzle is easy. Generating one with exactly one solution reachable by deduction alone is not — and an LLM cannot be trusted to check its own work here. Early cases were technically solvable but needed a guess somewhere, which is the cardinal sin of the genre. So every case now goes through a brute-force constraint solver before it publishes. If the solver finds two consistent worlds, or can only reach the answer by trial, the case is thrown away and regenerated.
The top tier is 5 suspects x 5 rooms x 5 weapons plus a fifth arrival-time axis, with some clues deliberately lying to you. That's an 8.6 million world search space, so verifying a case fast enough to publish one daily took some real work. Claude wrote that solver, it could never have replaced it.
WHAT IT'S ACTUALLY LIKE
Three things I'd tell anyone starting something this size:
- It is confidently wrong in ways that read as correct. The code looks right, reviews fine, and is subtly broken. You catch it by looking at the finished output, not at the code. Reviewing what it wrote is nowhere near enough on its own.
- It will tell you something is verified when it has only checked the easy half. If you don't insist on testing the actual thing, the real page, the real send, the real render, it will happily confirm a proxy for it and call that proof.
- It has no sense of what matters. Left alone it will polish something nobody will ever see while a real problem sits untouched. That judgement is still entirely on me, and it's most of the job now.
None of that is a complaint. Three months of evenings has produced something I'd never have got to alone. But the ceiling isn't the model, it's how well you can say what you actually want.
THE BIT THAT BELONGS IN THIS SECTION SPECIFICALLY
The single biggest source of visitors is now ChatGPT. Ahead of Reddit, and by a distance the most engaged traffic on the site at roughly 4 pages a visit. AI built the thing and AI is now what brings people to it, which I did not plan for and still find slightly funny.
Three months in, 300-odd cases published and a couple of hundred registered players. Happy to go into any of it; the generation and verification side is the genuinely interesting part and I don't see it discussed much.
Give it a try!
https://deduce.studio
I've been building Deduce — a daily deduction puzzle. Cluedo-style logic grid: work out who did it, where, and with what, from a set of clues. New case every day across five difficulty tiers. Free, no signup, no ads, nothing to buy.
It's written almost entirely by Claude, in this pattern: a full spec first (functional, technical, build plan), then three sub-agents run against that spec in parallel — one on the world database and API, one on the frontend and print layouts, one on the puzzle engine. The spec doing the coordinating is the whole trick. Agents working off a chat history instead of a document hand you three incompatible halves.
Stack is deliberately boring: React + Vite, Express, Postgres with raw pg and no ORM. Playwright renders each case to a print-ready PDF for a paperback format. The whole thing runs in a container on a mini-PC in my study behind Cloudflare, so hosting is R0/month.
THE PART AI COULDN'T DO
Generating a puzzle is easy. Generating one with exactly one solution reachable by deduction alone is not — and an LLM cannot be trusted to check its own work here. Early cases were technically solvable but needed a guess somewhere, which is the cardinal sin of the genre. So every case now goes through a brute-force constraint solver before it publishes. If the solver finds two consistent worlds, or can only reach the answer by trial, the case is thrown away and regenerated.
The top tier is 5 suspects x 5 rooms x 5 weapons plus a fifth arrival-time axis, with some clues deliberately lying to you. That's an 8.6 million world search space, so verifying a case fast enough to publish one daily took some real work. Claude wrote that solver, it could never have replaced it.
WHAT IT'S ACTUALLY LIKE
Three things I'd tell anyone starting something this size:
- It is confidently wrong in ways that read as correct. The code looks right, reviews fine, and is subtly broken. You catch it by looking at the finished output, not at the code. Reviewing what it wrote is nowhere near enough on its own.
- It will tell you something is verified when it has only checked the easy half. If you don't insist on testing the actual thing, the real page, the real send, the real render, it will happily confirm a proxy for it and call that proof.
- It has no sense of what matters. Left alone it will polish something nobody will ever see while a real problem sits untouched. That judgement is still entirely on me, and it's most of the job now.
None of that is a complaint. Three months of evenings has produced something I'd never have got to alone. But the ceiling isn't the model, it's how well you can say what you actually want.
THE BIT THAT BELONGS IN THIS SECTION SPECIFICALLY
The single biggest source of visitors is now ChatGPT. Ahead of Reddit, and by a distance the most engaged traffic on the site at roughly 4 pages a visit. AI built the thing and AI is now what brings people to it, which I did not plan for and still find slightly funny.
Three months in, 300-odd cases published and a couple of hundred registered players. Happy to go into any of it; the generation and verification side is the genuinely interesting part and I don't see it discussed much.
Give it a try!
https://deduce.studio