Good AI Bad AI

Smoketest is live

This baseline confirms we can ship quickly and iterate into the first playable puzzle level.

Checks

  • SvelteKit app booted
  • Cloudflare Worker adapter configured
  • Wrangler deployment pipeline ready
  • AI context/style/level scaffolding added

First playable level: Navigation

  • One AI action every second
  • One player reward/punish action per AI turn
  • 0.25s move burst + 0.75s wait cadence
  • Single-agent navigation from center to escape
  • Hidden policy weights + score that decays over time

Play now at /levels/navigation. See ai/levels/level-01-navigation.md for loop details.

Next layer: Dodging

scaffolded Dodging adds hazards on top of the navigation pacing.

Prototype route: /levels/dodging

Scaffolded styles

  • silly-friction: Readable objectives with awkward, funny execution.
  • toy-lab: Simple training-lab framing with visible state and feedback.

Style source of truth: ai/styles/style-profiles.yaml

Level manifest

  • shipped Smoketest — Validate runtime and deployment path.
  • playable Navigation — Simple escape loop with one AI action and one feedback action per second.
  • scaffolded Dodging — Advanced variant adding falling hazards on top of Navigation pacing.

Current focus: Navigation (ai/levels/level-01-navigation.md)