aiw-orchestrationlisted

Chooses the execution shape for a task — run it solo in the main loop, fan out read-only scouts, run an orchestrator over builders and reviewers, or relay a chain too large for one context window from agent to agent — gives its own sealed owner to a step the built system runs through a model on every use, and routes model capability and effort per role once a shape is chosen. Use this skill to decide whether to split work, before deciding rather than after: whenever a task spans many files or areas, its inputs exceed one context window, it looks like a batch of similar mechanical edits, its edits would collide in one tree, or there is any impulse to hand implementation to another agent, and whenever the deliverable includes a step a model will run on every use of the built system. Use it before handing writing work to any agent other than yourself. Deciding not to split is one of its answers and the more common one. It owns the gate that keeps multi-agent shapes off ordinary work, the per-role capability and
philippe-ths/ai-coding-workflow · ★ 0 · AI & Automation · score 66

Install: claude install-skill philippe-ths/ai-coding-workflow

# Orchestration Read this file when the work looks like it might split, before handing any writing work to an agent other than yourself, and when the deliverable includes a step a model will run on every use of the built system. Read it to decide whether to split, not once you have decided. ## The Default Is Solo One agent in the main loop is the default shape, and staying there needs no justification. Fanning out does. Before considering any multi-agent shape, spend the cheaper controls first, in this order: 1. **Fix context.** A miss caused by a vague brief, a missing file, or an unavailable tool is not fixed by adding agents. 2. **Adjust effort.** Effort governs how much the agent reads, runs, verifies, and persists through. A task that failed because steps were skipped needs more thoroughness, not more agents. 3. **Adjust capability.** A task that inspected the right evidence, genuinely tried, and still reached the wrong conclusion needs a stronger model. 4. **Only then, change the shape.** Skipping to step 4 is the common failure. A multi-agent run spends several times the tokens of a single agent, so the task has to be worth the multiplier before the shape is even a candidate. ## The Gate Fan out only when the work partitions into pieces that are genuinely independent — no piece needs another's output to start — **and** at least one of these holds: - The inputs exceed what one context window can hold. - The pieces are mechanical and fully specified, so a cheape