Metadata Fields

Use these fields to keep skill metadata consistent and easy to route.

Field	Purpose
id	Unique identifier (recommended rr-xxxx format).
name	Human-readable skill name shown in reviews.
description	Short summary of what the skill checks.
category	Primary routing key: core, upstream, midstream, or downstream. New skills should use this.
phase	Backward-compatibility alias for category. New skills should use category. Allowed values are upstream, midstream, downstream (core is only for category).
applyTo	Glob patterns for files the skill should inspect.
tags	Optional classifiers (for example, security or performance).
severity	Optional impact level: info, minor, major, critical.
inputContext	Required inputs the skill expects (diff / fullFile / tests / adr / commitMessage / repoConfig / reviewSelf / reviewExternal / findingsPool / prDescription).
outputKind	Output categories produced (findings / summary / actions / tests / metrics / questions / review-audit). Defaults to findings if omitted.
modelHint	Model selection hint: cheap / balanced / high-accuracy.
dependencies	Required tools/resources (for example code_search, test_runner, adr_lookup, repo_metadata, coverage_report, tracing, or custom:*).
priority	Optional ordering hint (lower runs earlier) for deterministic execution when no planner is used.

evaluationType (deterministic / heuristic / agentic) and deterministicGate are optional fields that declare which evaluation layer runs the skill. See Skill Schema for details.

Keep metadata in front matter so it can be parsed before the instructions run. All required fields must pass checks using /schemas/skill.schema.json.