add-standards-layer

Add a standards layer for $ARGUMENTS to this repository.

Where layers live

• .claude/rules/common/ — stack-agnostic rules. No frontmatter, loaded in every session. This is where non-negotiable invariants live, because they must apply while a file is being created and after /compact.
• skills/standards-<name>/ — language and framework layers. These load on demand when the description matches the work at hand, which also makes them distributable inside the plugin (.claude/rules/ is not a plugin component).
• .claude/rules/_template/ — blank canonical stubs to copy from.

Steps

1. Confirm there is a real delta. A layer exists only to say what common/ does not already say. If the guidance is general, it belongs in common/ instead — put it there and stop.

2. Create skills/standards-<name>/ with a SKILL.md and one file per topic where the layer has a genuine delta. Copy the stubs from .claude/rules/_template/; the canonical set is a menu, not a mandate:

   coding-style.md · patterns.md · security.md · testing.md · hooks.md

   Process topics (agents, code-review, development-workflow, git-workflow) are common/-only and must not be duplicated into a layer.

3. Write SKILL.md as a thin overview, under 500 lines, that says what the layer covers and points at the topic files. Its description is the trigger — name the concrete technologies and concerns so the layer loads when it should:

   ---
   name: standards-<name>
   description: <Name> coding standards — <the two or three things a reader would search for>.
     Use when writing or reviewing <name> code.
   ---

   For a framework layer, extend the language layer rather than common/, and say so in the body: > Extends [standards-python](../standards-python/SKILL.md).

4. Write only the delta. Every topic file opens with a > Extends [common/<topic>.md](...) line. If you catch yourself restating a rule from a layer above, delete it. If a rule turns out to be general across languages, move it up into common/ instead of repeating it.
5. Register the layer in:
   • CLAUDE.md — the map of the repository and the layer table.
   • .claude/rules/README.md — the taxonomy.
   • docs/ (and docs/ru/) — wherever layers are listed.
6. Wire tooling if the layer needs it — formatter/linter hooks in hooks/hooks.json and the matching entries in permissions.allow. Hooks must be self-guarding: no-op when the tool is not installed, and never block an edit.
7. Verify with /audit-config, then confirm the layer actually triggers: in a fresh session, a prompt about that language should load the skill without naming it.