Chief of Staff

1. 1 Decompose Claude Opus 4.8 940ms

   Two jobs, not one: groom the backlog, then summarise the delta. The second depends on the first.

2. 2 Route — 12ms

   backlog-grooming scores 86.5% — above the 80% floor. No worker in this workspace covers "brief me", so that step comes back to the user.

3. 3 Hand off Claude Haiku 4.5 380ms

   Passed the grooming output only — 3 issues touched, 1 escalation. The next step does not need the reasoning that produced it.

Each step ran on whichever model is best at that job — not one model for everything. The worker is the workflow.

• Read: the worker capability index, current eval scores and recent run traces
• Write: nothing directly — it only invokes already-deployed workers and records each hand-off
• Escalate: the user when no worker clears the score threshold for a step, or when a step has no owner
• Never: deploy, modify or configure a worker — it routes to existing ones only
• Never: route a step to a worker scoring below 80% on its own eval

• Routed step 1 to backlog-grooming (86.5%). Completed.
• Refused step 2 — no deployed worker covers briefing. Escalated instead of guessing.
• Handed off 180 tokens of output, not 4,200 tokens of transcript.

private rules were taught in one workspace and never leave it. general rules are true about the work, so every copy of this worker gets them.