---
title: "uno-tuner"
description: "Race UNO S1 prompts, tune wording from win-rate deltas"
canonical: https://agentpluginsdirectory.com/plugins/uno-tuner
last-updated: 2026-10-10
---

# uno-tuner
Race UNO S1 prompts, tune wording from win-rate deltas
- Slug: uno-tuner
- Publisher: hassenhamdi
- Repository: https://github.com/hassenhamdi/uno-s1-duel
- Manifest: powers/uno-tuner/plugin.json
- Version: 1.0.0
- Category (editorial): other
- Skills: 1 (duel)
- MCP servers: 0
- Stars: 0
- Repository created: 2026-10-04
- Repository last pushed: 2026-10-04
- Publisher type: User
- Listing: https://agentpluginsdirectory.com/plugins/uno-tuner
- Schema: https://agent-plugins.org/schemas/1.0.0/plugin.schema.json

## What uno-tuner does, in the publisher's words

A research program in miniature: pit StartLux-Decision-0.8B-Q8_0 (a typed decision model: no text generation, one forward pass per question) against real games, and find exactly where it wins, ties, and fails.

Main result: UNO, 2-player, RLCard engine: 22/30 vs random (p≈0.002), 17/30 vs a hand-coded tactician. A zero-shot 0.8B plays textbook UNO to a draw against written tactics, at ~35 to 85ms per move, with a human seat at a browser table.

▶ Watch: https://youtu.be/V2z4qdVKEok

- Narrate the state: full sentences, never tensors or tags. ("Snake head at (1,1). A snake body is south." beat "BODY-DIE" 0.71 v 0.31.)
- Consequence options: every choice states what happens if picked ("go south into your own neck and die", "sets your best color: red").

From the project README, punctuation lightly normalized. Full text: https://raw.githubusercontent.com/hassenhamdi/uno-s1-duel/HEAD/README.md