founder-os

Multi-agent primitives for Claude Code: consensus, debate, fanout research, skill building, and measured optimization.

TRIGGER when user explicitly invokes optimization with a numeric automatable metric — "optimize the lighthouse score", "shrink the bundle", "reduce build time", "cut p99 latency", "hill climb the eval accuracy", "/autoresearch", "make X faster" where X is measurable. Drops a working hill-climb harness into the project, generates project-specific mutations by reading the codebase, runs the loop. SKIP when there's no numeric metric, when changes need design/UX/copy judgment, or for one-off manual tuning. SKIP requests like "make this cleaner/better/smaller" that have no measurement command.

Fan-out/fan-in research. Spawns N researchers (Sonnet) in parallel to explore a question from different angles, then synthesizes findings with Opus.

Spawn Claude instances into a shared conversation room where they debate, disagree, and converge. Round-robin turns with parallel execution within each round.

Build flawless Claude Code skills. Studies existing skills as reference, ensures correct format, and pushes for genuine intelligence — skills that exploit something specific about how Claude works. SKIP for one-off scripts, prompts, or task helpers.

Stochastic multi-agent consensus. Spawns N agents with identical prompts to independently analyze a problem. Aggregates by consensus (mode), divergences (splits), and outliers (unique ideas).