Skill naming conventions

The compass- prefix marks a skill this project owns, and the second segment says which regulatory function it offers. Five shapes exhaust the catalogue.

compass-helm                                  the sticky mode
compass-principle-<rule>                      stance, cited not run
compass-method-<kind>                         a sequence of phases
compass-<domain>-<verb>[-<object>]            an action on a typed target
compass-act, compass-act-all, compass-export  dispatch

An action keeps the <domain>-<verb>[-<object>] form it has always had; the prefix is the only addition. principle- and method- are second segments rather than domains, because a principle names no artefact type and a method spans several.

• kebab-case throughout; the name is the directory under doc/llm/skills/, the deployed directory under .claude/skills/, and the name: frontmatter field — keep all three identical.
• Named after the intent, not the mechanism: the user thinks "I want to merge this PR", so the skill is compass-pr-merge — regardless of whether compass pr merge, gh, or anything else implements it.
• The object is omitted only when the verb is unambiguous within the domain (compass-pr-merge).

code- vs codegen-: manual versus automatic. Writing a unit test by hand is compass-code-add-tests; scaffolding an entity from a model is compass-codegen-add-entity.

Every entry above is preceded by compass-, so the domain is the second segment of an action's name: compass-pr-merge, compass-agile-start-task. Two further second segments are reserved, and they take no domain:

add versus start is the create-versus-pick-up distinction. Creating a story for backlog refinement must not create a branch; picking one up must.

• No roles. A skill is what we do, not who does it. Role skills are dissolved into their constituent actions (e.g. agile product owner → compass-agile-refine-backlog, compass-agile-plan-sprint).
• Reference-only skills, except principles. If the instructions contain no executable steps, it is usually knowledge wearing a skill costume: demote it to a knowledge doc and link it from the skills that act. The citation function is exempt. A principle is by definition a rule that is cited and never run, so requiring executable steps of it would forbid the whole function; it earns its place as a skill because a phase must be able to load it by name at the moment of judgement.
• Every skill declares its level. #+level: is mandatory. It is what lets a declared System attenuate the catalogue, which is the whole mechanism by which a session sees a set small enough to choose from well.
• One skill per runbook. Runbooks are skills' procedural twins; a generic runbook dispatcher hides the actions from discovery.
• Check before adding. A new skill name must parse against the two registers, and the skills inventory must be checked for an existing skill covering the same action. Extending a register is a deliberate act recorded in the decision log below.

#+level: carries one of these values. The level sets the default slice of the catalogue a session sees; the full treatment is in Systems and levels.

Attenuation is advisory at S1, S2, S3, S4 and S5: a default small enough to choose from well, which an agent may reach outside by saying so. It is binding at S3*, because an audit channel that can be talked into editing what it observes is not an audit channel.

A name is only decidable once the artefact type is. This is the placement procedure from Regulatory Functions, reproduced where an author naming something will meet it. Take the first match.

1. A durable fact about the domain goes to a knowledge document. If it is the entry point to a cluster, a structure note.
2. An account of how one operation is performed here, with a runnable script, goes to a recipe. This is also where a method's missing effector goes.
3. A correction that must not be repeated goes to a memory. Keep these scarce: each is paid for by every session.
4. A rule that is cited rather than run goes to a principle.
5. Intended behaviour that should be checkable goes to a specification.
6. A sequence of phases ending in a verified state, not tied to one artefact type, goes to a method.
7. A change to a generated artefact goes to its model or its template, never to the output.
8. An operation on one artefact type goes to an action, at the pairing of type and verb. The action holds the judgement; the recipe holds the command.
9. A contested judgement goes to deliberation, and its verdict is recorded against the work.
10. A role a session or a delegate acts as goes to a System, tagged by level, stating what it must read first and what it may do.
11. The argument for a change, spanning many tasks, goes to a design in the story's own folder rather than to the durable tree.