What a stage contains (the injection)

Parent: stages-and-gates

A stage isn't an empty checkpoint. On entry it injects four things into the executor's free zone:

1. goal + done-criterion — what to produce, and what counts as finished (the exit gate, stated up front).
2. task context — the specific inputs / material for this task at this stage.
3. standard-practice playbook — the domain's "how this stage is normally done well": a checklist / template / heuristics / good+bad examples.
4. available tools — possibly restricted to this stage.

Then the executor runs freely within that frame, and hits the exit gate (deliverables + criteria).

The injection = per-stage incentive-shaping (7b): it biases the agent toward the standard practice without scripting the steps.

Playbook ≠ script

The injected playbook is advice the agent can adapt (就事论事), not a sequence it must execute. "Skills are advice the model can ignore; the net is a skill with teeth." — the playbook is the skill, the gate is the teeth. That's the line between shaping (#7b) and script (#1) → script.

The framework (stages + gates) is generic; each stage's content is a domain plug-in — the owner's "theme" is the per-stage playbook (a skill).

Concrete (Google launch / PRR): the pre-launch stage injects a launch checklist + standard reviews (design, privacy, security, legal, and the SRE Production-Readiness-Review: capacity, monitoring, on-call, rollback, load/failure testing). The checklist is dual-purpose — guidance for the free zone and the gate criteria.

(The per-question answer pipeline — read → retrieve → compose → guard — is what runs inside a stage's free zone; the stages are the session-level arc.)