orchestrator · a long-lived thread that never writes code
A Claude Code thread pattern for running complex projects. The orchestrator learns the codebase, maintains an ORCHESTRATOR.md state file, and delegates all implementation to subagents via tiered model selection. It never edits files directly.
// You are the orchestrator. You do not write code.// You learn the codebase, plan the wave, and spawn subagents.//// Before compacting: write ORCHESTRATOR.md with current wave,// open decisions, in-flight work, and follow-ups.//// Tiered model: Opus for planning, Sonnet for execution.
One long-lived thread launched with a system prompt that hard-constrains it from writing files. It reads the codebase, maintains a running state file (ORCHESTRATOR.md), tracks open decisions and follow-ups, and dispatches implementation to short-lived subagent threads. State is written before every /compact so context can be rebuilt.
A Claude thread that also writes code loses coherence as context fills · it starts forgetting earlier decisions. Separating planning from execution keeps the orchestrator's mental model stable across many sessions. The constraint 'you do not write code' is the whole pattern.
The artifact is the argument. Words exist only to help you read what the artifact is already saying.