Playbook

What’s it for?

A coding agent can rediscover the same release, recovery, or diagnostic procedure from one task to the next. The commands and decisions live in a transcript, so the next run starts from scratch.

Playbook starts by finding every command, skill, and instruction that claims to own the work. It reuses an authority when its contract is right, repairs it when incomplete, and removes it when stale or duplicated. New infrastructure is the last option; a safe one-off adds nothing.

A concrete case: Stringproof

Apple .strings files can still parse after translation while losing a key or placeholder, gaining prompt residue, or mixing scripts. Given source and target catalogs, those checks are deterministic. Under Playbook, they belong in a CLI—not another skill.

$ stringproof check fixtures/corrupted --source-locale en
ja.lproj/Localizable.strings:1 [key.missing] ja settings.open: key is present in source catalog but missing from target
ja.lproj/Localizable.strings:2 [placeholder.missing] ja welcome.user: missing placeholder(s): %@
ja.lproj/Localizable.strings:3 [placeholder.added] ja items.count: added placeholder(s): {total}
… 31 more findings

That decision produces something testable: stringproof check reads a directory, reports stable findings, and exits 0, 1, or 2. Stringproof documents the command; the repository contains its implementation and fixtures ↗

What Playbook decides

Work qualifies after it has recurred or when another operator must replay or recover it exactly. It also qualifies when improvisation is risky, the same judgment recurs across tasks, or an external service needs a reusable interface. A missing wrapper or possible future use is not enough.

Check it against the task

The agent must finish the current task through the authority it leaves behind. Doing the real work tests a reused, repaired, or new interface against its inputs and failures. If the task still depends on remembering the old sequence or a removed rule, the work was not captured.

Playbook handles this decision only. It uses commands for deterministic work and skills for recurring judgment. It rejects parallel wrappers and removes authorities that no longer match the workflow.

Install

Tell your coding agent:

Install the Playbook skill using https://workcell-137.pages.dev/playbook/llms.txt

Then use it on work that has already qualified:

Use $playbook to decide what should persist from this task, then complete the task through it.

The skill does not expand the task or its permissions. Read the skill on GitHub ↗