biome: Skill for Claude Code

Lint Rule Development

Follow the current analyzer architecture in crates/biome_analyze/CONTRIBUTING.md. Read only the sections relevant to the rule rather than loading the entire guide by default.

Workflow

1. Find two or three current rules in the same language and group with a similar query or action.
2. Generate the matching lint-rule or assist scaffolding when adding analyzer behavior.
3. Implement the narrowest query and state needed to decide whether to signal.
4. Add a diagnostic and, when safe, an action.
5. Add focused valid and invalid fixtures, then inspect snapshots.
6. Run required analyzer codegen and the narrowest tests.

Scaffolding commands:

just new-js-lintrule useMyRule
just new-css-lintrule useMyRule
just new-json-lintrule useMyRule
just new-graphql-lintrule useMyRule

New lint rules start in nursery. They are patch changes targeting main, because nursery rules do not follow normal feature versioning. Load changeset for the release entry.

For a new assist, use the language's new-*-assistrule recipe, for example:

just new-js-assistrule useMyAction
just new-json-assistrule useMyAction

The generator places assists under src/assist/source/; they do not use lint groups or the nursery policy. A new assist is a user-facing feature and normally requires a minor changeset targeting next. A bug fix to an existing assist follows normal bug-fix policy. Check justfile for the languages with assist scaffolding.

Query and State

Choose the least expensive query that answers the rule:

• Ast<Node> for syntax-local checks;
• Semantic<Node> when bindings, references, scopes, or globals are required;
• a service query only when the fact is owned by that service;
• type inference only when syntax and the semantic model cannot answer the question.

The analyzer guide's query and service sections are canonical for available APIs.

run() should decide whether to emit a signal. Keep action-only work in action() so it is not performed for every candidate node.