agentic-delivery-playbook

Agentic Delivery Playbook for Pi

Use this skill to make coding-agent work smaller, safer, and easier to review.

This skill is the policy layer for Pi delivery. It should coordinate Pi primitives instead of replacing them:

agentic-delivery-playbook = triage, scope, approval, evidence policy
pi-goal-x                 = durable lifecycle for broad/long-running work
pi-subagents              = scoped implementation and review lanes
run artifacts             = evidence trail only when useful

This is a value gate, not ceremony. If applying the playbook will not change the next action, skip it and implement directly.

The value test

Use the playbook only when it will do at least one of these things:

• narrow a vague request into a concrete goal and non-goals
• stop risky work before implementation for human approval
• define acceptance criteria and validation evidence before coding
• prevent model/agent routing claims that are not actually true
• split broad work into bounded implementation tasks
• create reviewable closeout evidence for a PR or future agent run
• coordinate a Pi goal, worker, reviewer, or planner without losing parent ownership

If none of those apply, say Mode: Direct, make the smallest safe edit, validate, and report evidence.

What this skill is not

• Not a mandatory spec-first workflow for every task.
• Not a requirement to create run artifacts for small edits.
• Not a hard-coded model benchmark or vendor policy.
• Not permission to launch a subagent swarm.
• Not a replacement for Pi goal lifecycle or subagent orchestration.
• Not evidence that a model was used unless Pi actually routed that model and the run recorded it.

Native Pi delivery shape

Use the smallest combination of Pi primitives that makes the work safer:

Mode	Pi goal	Subagents	Artifacts	Default owner
Direct	No	No	No	parent implements
Lightweight	Usually no	Optional focused reviewer or worker	Compact checklist/spec only if useful	parent, or worker when separation helps
Full	Recommended, especially if long-running/broad	Yes by default: worker for implementation, reviewer for QA; planner/scout when useful	Yes	parent orchestrates, worker writes
User says “goal”, “work until done”, or asks for durable progress	Yes	As needed	Yes	goal lifecycle owns stop/complete