AlfredStar

Each cycle the agent reads StarCraft II's current state: minerals, vespene, supply, army composition, buildings, and whatever the enemy has shown so far. The live match view on the practice page is reading the same snapshot.

Given what it sees, the agent decides what to do next: which units to train, which buildings to add, whether to hold or push. Each decision lands as a DIRECTIVE entry in the practice log so you can watch the plan evolve as the match unfolds.

A deterministic engine carries the directive into the game — queueing builds when the economy can afford them, producing units, moving the army. You see its output as the live snapshot updates and as the match progresses.

A practice is one turn of the outer loop:

1. AlfredStar loads your playbook (and any focus you set) into the agent's context.
2. The agent plays one SC2 match against the chosen opponent — the inner loop above runs cycle after cycle until the match ends.
3. A separate reflection step reads the match's logs and writes a short summary: what worked, what surprised the agent, where the playbook fell short.
4. You read the reflection, refine the playbook, and queue the next practice.

A match is bounded — the agent's decisions during it are fixed by the playbook it loaded at the start. Improvement happens between practices: you're the loop, the reflection is the signal, the playbook is the delta you push into the agent for next time.

You write rules — short instructions — into a playbook. The agent loads the playbook into its context each match. You can also set an optional focus when starting a practice to nudge that match toward a particular emphasis.

AlfredStar runs practices with your Claude API key — bring your own on the Account screen after signing in.