A userscript for krillion.io that asks an AI for rare answers to the prompt on your screen. It learns from your real results, so its suggestions get better the more you play.
Not affiliated with Krillion or its creator. Krillion is a one-person project: if you enjoy it, consider supporting it.
⬇️ Install (one click) (you need a userscript manager first; see Install).
Krillion gives you prompts like "name a cheese". Obvious answers (cheddar) are worth almost nothing, and rare ones (taleggio) score big. Krillion Helper adds a small panel to the page that:
- Suggests rare answers as soon as a prompt appears, streamed in one by one, rarest first, with a 1–5 ★ rarity guess.
- Checks them before you see them:
- Letter and length rules in the prompt ("starting with B", "5-letter", "ending in stan", …) are checked locally.
- An optional second AI pass throws out made-up or wrong answers. Answers that pass get a ✔.
- Learns your real results. After each guess it reads the tier the game showed you (plankton → schooler → rare → deep cut → krillion, or 🤡 too clever) and replaces the AI's guess with the truth. Answers the game rejected are hidden for good, and the AI is told about them next time.
- Saves everything locally, so a prompt you've seen before shows its answers instantly, with no AI request.
- Puffer Party aware: it watches which answers have already been said, hides them from your suggestions (repeats get rejected), and tells the AI to skip them.
- Install a userscript manager:
- Tampermonkey (Chrome, Edge, Firefox, Safari; recommended, and the only one that streams answers live).
- Or Violentmonkey.
- In Chrome/Edge, also turn on "Allow User Scripts" in Tampermonkey's extension details, or Developer mode at
chrome://extensions.
- Click Install Krillion Helper, then press Install.
- Open krillion.io. The panel appears bottom-right. Click ⚙, pick an AI provider, and enter its API key (see below).
It updates itself automatically whenever a new version is pushed here.
Pick a provider in ⚙ → AI, enter its key with the API key button, then press test to check it works. list models shows every model your key can use.
| Provider | Notes | Get a key |
|---|---|---|
| DeepSeek | Cheap and good; the default | platform.deepseek.com |
| Claude (Anthropic) | Opus is the smartest; Haiku is fast and cheap | platform.claude.com |
| OpenAI | GPT models | platform.openai.com |
| Google Gemini | Has a free tier | aistudio.google.com |
| Groq | Extremely fast; has a free tier | console.groq.com |
| OpenRouter | One key, hundreds of models | openrouter.ai |
| Mistral / xAI Grok | mistral · x.ai | |
| Ollama (local) | Free, runs on your PC, no key | ollama.com |
| LM Studio (local) | Free, runs on your PC, no key | lmstudio.ai |
| Custom | Any OpenAI-compatible server URL |
Local models (Ollama): install Ollama, run ollama pull llama3.1:8b, and pick Ollama in ⚙. If you see "blocked by the local server", restart Ollama with the environment variable OLLAMA_ORIGINS=*. Small local models make more mistakes, so keep double-check on and use depth: safe.
Cost: each prompt is one short request, plus one more if double-check is on. On DeepSeek, Haiku, Gemini Flash or Groq, a whole day of dives costs a fraction of a cent. Saved prompts cost nothing.
- Click a suggestion to put it in the answer box, then press Enter. Or turn on picking also submits in ⚙.
- ↻ get new answers asks for more, never repeating ones you already have.
- Drag the panel by its header to move it. Click the header to collapse it.
| Shortcut | Does |
|---|---|
| Tab / Shift+Tab (in the answer box) | Cycle through suggestions |
| Alt+1 … Alt+9 | Pick suggestion 1–9 (on Linux Chrome these switch tabs, so use Tab there) |
| Alt+R | Get new answers |
| Alt+A | Ask for the current prompt |
| Alt+H | Hide or show the panel |
What the labels mean:
| Label | Meaning |
|---|---|
| ★★★★ | The AI's rarity guess (1–5) |
| ✔ ★★★ | Passed the double-check |
| ✓ DEEP CUT | The game's real tier for that answer |
| ✗ not accepted | The game rejected it (hidden by default) |
AI
- Provider, model and server URL (local or custom only).
- API key, list models, test.
- Thinking effort (Claude only).
- Creativity, and whether to send it at all.
- Timeout.
Answers
- Answers per request.
- Depth:
- safe: more hits, lower rarity.
- balanced: the default.
- deep: riskier, higher scores.
- Minimum ★ to show, double-check, and auto-ask on new prompts.
- Fetch more automatically when you're running low.
- Picking also submits.
- Hide answers already said in a party, hide past misses.
- Extra instructions: free text added to the AI's instructions, e.g. "prefer British spellings".
Auto-answer (solo modes only)
- Submits the best suggestion on a timer: a wait, a random extra wait, and a wait after a miss. The panel shows a live countdown.
- If an answer is rejected it tries the next one, and asks the AI for more if it runs out.
- It never runs in Puffer Party, where you're playing live against real people.
Look
- Corner, width, text size, opacity, shortcut hints.
Saved answers
- Forget the current prompt, or clear everything.
- Export / import to move what you've learned between computers. Import merges files and keeps real results over guesses.
- Reset position, reset settings.
- API keys are stored only inside your userscript manager on your own computer. They're never written into the script or sent anywhere except the provider you picked.
- What's sent: only the prompt text, plus the answers already tried for it, goes to your chosen AI provider. With a local model, nothing leaves your computer.
- Network permission: the script lists the AI providers it talks to. It also has
@connect *so local and custom servers work, and Tampermonkey asks your permission before it contacts any server that isn't on that list. - What it never touches: the script doesn't call Krillion's servers, doesn't read hidden answer lists, and doesn't unlock paid packs. It only reads what's already shown on your screen.
This is a study and practice aid. A few things to keep in mind:
- Daily scores are compared against everyone else's. Scores you post with AI help aren't really yours, so think about that before sharing them or comparing with friends.
- Auto-answer is deliberately disabled in Puffer Party. Please don't modify it to play live games against people.
- The developer may consider AI help against the spirit of the game. Use it at your own risk.
| Problem | Fix |
|---|---|
| Panel says "waiting for a prompt…" forever | Refresh. Check the script is enabled in Tampermonkey. Krillion may have changed its page layout, so open an issue here |
| "API key rejected" / "out of credit" | Re-enter the key (⚙ → API key) and check your balance with the provider |
| "model or URL not found" | Press list models and pick one from the list |
| Lots of answers rejected by the game | Use depth: safe, keep double-check on, or try a smarter model |
| Answers appear all at once instead of streaming | Live streaming needs Tampermonkey; other managers deliver the whole reply at the end |
| Ollama "blocked" | Start Ollama with OLLAMA_ORIGINS=* |
- The script finds the current prompt from the round card on the page and reads your results from the result card the game draws after each guess.
- Rejected guesses are read from the "not in the list" message, and Puffer Party answers from the players' speech bubbles.
- Requests go through Tampermonkey's
GM_xmlhttpRequest, so any provider works without CORS problems. - Everything except Claude uses the OpenAI-style
/chat/completionsAPI, and Claude uses the Anthropic Messages API. - Everything it learns is stored with
GM_setValue.