orchestratelisted

# Orchestrate — parallel work packages with async human decisions Turn one parent task into small work packages executed by parallel subagents, while every decision only the human can make becomes a plain-language question document with a stated default — so the human answers on their own schedule and nothing ever stalls waiting for them. Three roles: **the orchestrator** (you — decompose, launch, verify, keep shared state), **worker agents** (each produces exactly one output file), **the human** (answers question docs asynchronously, does the steps only they can do). ## The five invariants Everything else adapts to the project; these do not. 1. **One package = one agent = one output file.** A worker creates or edits exactly one file and touches nothing else. This is what makes N agents safe in parallel — no merge conflicts, no racing writes on shared files (a real hazard on Drive/Dropbox-synced or concurrently-open repos) — and it makes verification trivial: read the file, check the done-when. 2. **Packages are small enough to finish without a human.** Size each package so no human decision is needed *mid-package*. If a decision might arise, it becomes a question doc and the package proceeds on a default. A package that must pause for an answer was cut wrong — split it. 3. **Never block.** Workers and orchestrator alike: when a human-only decision or an underivable fact appears, write a question doc (see below), state a default, continue on th