Agent Skills: Claude Skills and Agent Code Review Patterns

Claude Skills and Agent Code Review Patterns

Stack-specific rules loaded by dh:code-reviewer when SKILL.md, agent frontmatter files, or plugin.json files are detected.

Frontmatter Validity

• name must match the containing directory name exactly — mismatch causes lookup failures
• name must satisfy ^[a-z][a-z0-9-]*$ — lowercase letters, digits, hyphens only
• description must be a single-line string — no YAML multiline indicators (>-, |-, >, |)
• tools field (in agents) must be a comma-separated string, not a YAML array — Read, Grep, Glob not [Read, Grep, Glob]
• model must be one of sonnet, opus, haiku, or inherit — no version strings, no full model IDs
• Run uvx skilllint@latest check --fix <path> after every frontmatter edit — auto-fixes YAML formatting issues including quoting

Description Quality

• The description is the primary routing mechanism when Claude selects from 100+ available skills — it must contain specific trigger keywords
• First 1024 characters are most important — front-load trigger conditions and activation signals
• Preferred trigger openers: "Use when", "Activates on", "Triggers on", "Loaded automatically when"
• Descriptions must be third-person ("Generates commit messages" not "I can help you generate")
• Vague descriptions produce missed triggers — "Helps with Python stuff" is a blocking finding

Token Budget

• Skills must not include information Claude already has from training (general language patterns, well-known APIs, common idioms)
• Include only project-specific context, conventions, and domain knowledge that Claude cannot infer
• Skills over 4000 tokens should be flagged as oversized (warn); over 6400 tokens are blocking
• Inline examples are valuable — prefer 2-3 focused examples over exhaustive lists

Scope of Tools

• allowed-tools in agent frontmatter must be scoped to the minimum required for the agent's task
• Agents that only review code must not have Write or Edit in their tools list
• Verification-only agents must not have destructive tools

Non-Invocable Skills

• user-invocable: false must be set for skills intended only for agent context injection, not direct user activation
• disable-model-invocation: true must be set for skills that execute shell commands or other side-effecting operations that should not be run as prompts

Agent Contracts

• All agents that return results to an orchestrator must include STATUS: DONE or STATUS: BLOCKED as the first line of their terminal response
• STATUS: BLOCKED must include a NEEDED section listing specific missing inputs
• STATUS: DONE must include an ARTIFACTS section listing what was produced and where
• Agents must not guess at missing inputs — BLOCKED is always correct when information is unclear

File Reference Standards

• All file references in SKILL.md use markdown link syntax: [text](./path/to/file.md) with ./ prefix
• Skills in subdirectories under skills/ silently fail to register — all skill directories must be directly under skills/ (one level deep only)
• Cross-skill references use activation syntax — "For X, activate the /plugin:skill-name skill" — never backtick paths to other skills' internal files
• No docs/ path references at runtime — developer documentation must not appear in agent-facing skill content

Code Fence Requirements

• All code fences must have a language specifier — bare triple backticks are a blocking finding
• Nested code blocks use 4 outer backticks and 3 inner backticks
• Markdown links use ./ relative paths, not absolute paths or backtick references

Plugin.json Auto-Discovery

• agents, skills, and commands keys in plugin.json are only for non-default locations — omit them when all components are in their default directories
• Declaring a subset of agents in the agents key overrides auto-discovery completely — unlisted agents become invisible
• If any component path is declared in plugin.json, every component in that category must be listed

Anti-Patterns

# WRONG: multiline description
description: >-
  This skill helps with TypeScript code review
  when the user asks for a review.

# RIGHT: single-line description
description: TypeScript-specific code review patterns covering strict mode, ESM, type safety, and branded types. Loaded automatically when reviewing TypeScript code.

# RIGHT: single-line description — run `uvx skilllint@latest check --fix <file>` to auto-fix any YAML quoting needs
description: Code review for TypeScript — strict mode, type safety, branded types, and ESM patterns.

<!-- WRONG: cross-skill backtick path reference -->
See `plugins/plugin-creator/skills/prompt-optimization/SKILL.md` for optimization patterns.

<!-- RIGHT: activation syntax -->
For prompt optimization, activate the `/plugin-creator:prompt-optimization` skill.