Playbook
How a task is routed to a playbook, and the principles each route applies.
/playbook matches a task to a playbook and copies its steps into the todo list verbatim. The playbook skill owns the routes.
Playbooks
- Backlog item. Probes the repo to re-derive current state from the plan's intent, then routes below through `bin/skills brief`, which prints the routed playbook and Handing back, so neither file is opened.
- Investigation. How does X work, are we sure about Z.
- Bug fix. A defect, runtime evidence.
- Perf issue. Measured slowness, a baseline.
- Feature. You own the design and the code. Plan, write, review, verify.
- Refactoring. Rename, extract, dedupe, move.
- Prototype. A throwaway sketch to settle an observable fork.
- Babysit. You own the merge frontier. Declare a mode, clear one PR at a time, stop where the human's call begins.
- Session pickup. Resume prior work.
- Pause safely. Suspend cleanly.
- Authoring or modifying a skill. You own the skill's voice.
- Handing back. Every playbook's last step: report, evidence, one commit message, stop.
Backlog item is what /plans do runs, and Handing back ends every playbook.
Principles
The first todo item reads the playbook's Principles section in full. The reply names each principle that shaped a decision and the choice it changed.
Each principle is a standalone principle-<slug> skill, slash addressable and read in full before it is applied. The playbook indexes the principles with a trigger for each. They set disable-model-invocation: true and carry a Codex agents/openai.yaml, so on Claude Code and Codex they never enter the model's skill listing and cost nothing per session. The Skills catalog lists them with the other skills.
Destination Gate
The destination gate runs first, before the probe. Under /plans do, it runs before the branch is cut. The plans skill's lint checks that the plan has a surface: value. The agent then checks whether the plan names exactly one destination and who can see it.
If the plan does not name exactly one surface, or public versus private is inferable from more than one place, the agent stops and asks. This overrides principle-never-block-on-the-human by name. The Backlog item playbook states the gate in full.
Outside /plans do the same rule holds for any task the playbook routes: a destination question is the one fork the agent asks about instead of trying something.
One incident put this gate first, and it happened while implementing a plan, not writing one. A plan named a tab, which read as the public profile, while a private directory of the same name sat beside it. The guess shipped a private breakdown to a public page. A prototype cannot settle which destination was meant, and neither can the nearest matching folder.