The workflow half of agent-driven development: how to orchestrate rounds, what a coder is held to, how a review is scored, how a disagreement gets settled, and how each round's lessons become the next round's rules.
None of it is about a particular agent. These skills were extracted from
pi-pilot after measuring what they actually contain: of
1,718 lines across eight skills, only pi-pilot's own 149 mention pi at all. The other seven never
mention it once. The pi- prefix claimed a coupling that did not exist, so it is gone.
The harness drivers live in their own repositories and are interchangeable:
Both load the workspace's AGENTS.md, so the project's own standard reaches the agent and a
brief can describe the task instead of restating the rules.
orchestrator |
decides. Plans rounds, scores them out of 50, ships or sends back, and distils what each round taught. Includes CHECKLIST.md — read before every launch, every verdict, every landing |
coder |
the standard a coder is held to: demonstrate, do not assert |
reviewer |
how to review so that a finding survives contact with the code |
adjudicator |
merges parallel reviews, settles disagreements by measuring rather than averaging |
prescription-verifier |
walks a prescription against the actual call sites before it is accepted |
distil |
turns a finished round into rules, so the next round starts further along |
instruments |
how to tell a working measurement from one that reports success while measuring nothing |
A check that can only ever pass has not been shown to work. Break the property deliberately,
watch the check go red for that reason, restore it, and watch it go green. Most of
CHECKLIST.md is the record of times that was skipped and what it cost.