Skip to main content
Every game starts from a default policy. Before any optimization loop can climb a leaderboard, that default must (a) exist, (b) run end-to-end without crashing or timing out, and (c) be structured so you can make one scoped change at a time. The goal is the smallest policy that already beats the bundled reference and is shaped for upgrades — not the cleverest policy.
1

Read the game mechanics

Answer the mechanics questions in Choose a game: observability, determinism, latency budget, players/asymmetry, score shape, protocol. Write durable findings into the game’s own notes.
2

Pick the architecture

Map mechanics to scripted / LLM / hybrid using the decision table. When unsure, seed scripted and let evidence justify cognition later.
3

Structure for long-horizon upgrades

Separate the brain from the transport, type the observation, emit a per-episode artifact, make every decision attributable, and keep one knob per concept. If the game is deterministic, build a local simulator.
4

Package as a Coworld player image

linux/amd64, importable package, correct --run argv. See Packaging. Enforce the robustness contract: a legal default action on any bad input, and an immediate scripted fallback for any LLM failure.
5

Verify locally before spending anything

Run coworld run-episode ... -n 1 -o runs/smoke and confirm scores are real (not a timeout), logs are clean (no tracebacks), and artifacts were written. See Verify locally.
6

Upload, verify, hand off

Verify the active player, read the live standings, upload, and confirm the run attribute is non-null on the new version. The seeded policy is now the base ref for the optimization loop.

​
Anti-patterns

  • Shipping a default that crashes or times out — a working simple policy always beats a clever broken one.
  • Uploading without the correct --run argv, or skipping the run-attribute check afterward.
  • Reaching for an LLM when the game is fully observable, deterministic, and latency-tight.
  • An LLM/hybrid policy with no deterministic fallback path.
  • A monolithic policy where brain and transport are tangled.
Source: adapted from Metta-AI/optimizer-agent skills/seed-a-new-policy/SKILL.md.