Co na to AI

Daily Czech guessing game in the spirit of Family Feud / Co na to Češi: „100 AI jsme se zeptali a teď hledáme osm nejčastějších odpovědí na otázku…“ Live at https://ai100.appwrite.network. Every day one survey question goes to 100 AI answerers (a different OpenRouter model per answerer, each playing a different Czech persona). Their answers are merged, and the player guesses the 8 most common. Five wrong guesses and the game is over.

Scoring: an answer is worth 10 points per AI that gave it, and clearing the board adds 200. Tapping a hidden answer offers two clues. "Odkrýt písmeno" reveals the next letter, up to 3, and never the whole word; each one takes a quarter of that answer's points, so a hinted answer is always worth more than a miss and less than finding it with fewer letters. "Kdo to řekl?" is free: it describes the personas who gave the answer ("hlavně důchodci z Moravy") and names the youngest and the oldest.

Stack: TanStack Start, React 19, Tailwind 4, shadcn/ui (radix). Node 22.18+ (the pipeline runs TypeScript directly with Node's type stripping).

Run the game

npm install
npm run dev        # http://localhost:3000
npm test           # matching, scoring, Prague day switch, stats, pipeline shapes
npm run build      # production build in dist/

Routes: / today's game, /hra/YYYY-MM-DD a past day, /finale this week's Finále (/finale/<Monday> for earlier weeks), /archiv the archive, /dotazani the 100 respondents (/obsazeni redirects there), /vlastni/<id> a custom game, /admin the admin for accounts with the admin label. /sitemap.xml lists the main pages and every published day, and robots.txt points to it.

A new puzzle starts at 00:00 Europe/Prague. Puzzle #1 is 2026-10-01 (LAUNCH_DATE in src/lib/day.ts). The server only returns puzzles whose date is on or before today in Prague, so the client never receives future boards. The answers in a payload are XOR'd with a per-day key and base64-encoded, so they aren't readable at a glance. That is obfuscation, not security.

The pipeline

echo 'OPENROUTER_API_KEY=sk-or-...' > .env

npm run pipeline                         # make sure today + the next 6 days have puzzles
npm run pipeline -- --days 14            # keep 14 days queued
npm run pipeline -- --date 2026-11-02    # one day
npm run pipeline -- --date 2026-11-02 --force   # redo a day
npm run pipeline -- --recluster 2026-11-02      # re-run only the judge on the saved answers
npm run pipeline -- --rejudge puzzles/2026-11-02.json   # re-cluster a published board from its stored answers
npm run pipeline -- --aliases puzzles/2026-11-02.json   # rewrite only the answer key (accepted guesses)

For each date the pipeline does seven things:

1. Takes the question scheduled in questions/YYYY-MM.json.
2. Asks the 100 answerers in data/answerers.json. Each is a model + persona pair. The prompt is in Czech, temperature 0.9, and asks for a single word, not one repeated from the question, as structured JSON when the model supports it, plain text otherwise. The persona only lightly colours the answer. A failing model is retried, then replaced by a spare from pipeline/lib/roster.ts. Reasoning effort is set from the catalogue: off where a model doesn't need it, the lowest level where it is mandatory.
3. Sends the raw answers to the judge model (JUDGE_MODEL, default anthropic/claude-opus-5.5). The judge merges them into single-word canonical answers: same thing, same root (zákon, zákoník), inflections, typos; a list answer counts for its most common item. Every answer must be used exactly once, otherwise the judge is asked again.
4. Asks the judge model to review the merged answers against the question and drops the non-answers: an answer that repeats the question or names what it asks about ("Bunda" for "Jaký kabát nebo bundu nosíte v zimě?"), the category itself ("Oblečení" for "Co si vezmete na sebe do posilovny?") and catch-alls like "Věci". Those votes count as invalid, and the dropped words stay wrong for players.
5. Runs the quality gate: at least 80 answers, the top answer at 45 % or less, at least 8 answers with 2 or more votes, the board holding at least 50 % of answers, nothing offensive. A question that fails is marked rejected, and a reserve question (questions/reserve-*.json) is tried for that date instead.
6. Writes the answer key. The judge model lists, for each board answer, every word a player might type for it: forms, diminutives, slang, brands, examples, and umbrella words ("alkohol" → Pivo). No word may fit two answers, and the names of answers outside the board stay wrong. On top of that, the client forgives diacritics, i/y, w/v, doubled letters and small typos.
7. Writes puzzles/YYYY-MM-DD.json (the board, the long tail, who said what, cost) and runs/YYYY-MM-DD--<question>.json (every raw response, attempt, latency and cost, plus the judge's output).

Each run prints the top 8, the top share, the cost and any gate flags.

.github/workflows/puzzles.yml tops the queue up every night and commits the new files. It needs the OPENROUTER_API_KEY repository secret.

Adding questions

Questions live in questions/YYYY-MM.json:

{ "id": "chata", "date": "2026-10-24", "text": "Co si berete s sebou na chatu?", "tags": ["chata"], "status": "approved" }

status is draft | approved | used | rejected; the pipeline only uses approved ones. The bank is dated through 2027-10-31 and has 50 undated reserves. A good question has an open but finite answer space. Avoid one obvious answer ("Jmenujte ovoce" gives 70 % jablko) and avoid hundreds of equally likely ones. Keep it under 70 characters.

The question bank

Questions come from questions/ (dated through 2027-10-31, plus reserves) and from the questions table, which refills itself:

This is the beginning of the README. Read the rest at the original. Read the full README ↗