Skills with Claude Code

Claude Code is a strong environment for skill development because it combines:

• local filesystem access
• shell/tool execution
• iterative editing in-repo
• agent/subagent workflows

CLAUDE.md vs Skills

Use both, but for different scopes:

• CLAUDE.md: always-loaded project memory (architecture, stack, global conventions)
• SKILL.md: task-specific procedural playbooks loaded when relevant

Use /init to bootstrap project CLAUDE.md, then keep repeatable workflows in .claude/skills/.

Practical Workflow

1. Add or install the skill collection you want to use (for example from anthropic/skills).
2. Verify available skills from inside Claude Code.
3. Trigger the skill with a clear task prompt.
4. Review generated file diffs and script outputs.
5. Refine SKILL.md description/body until activation and output quality are stable.

Typical Local Paths

• Project-scoped: .claude/skills/<skill-name>/SKILL.md
• User-scoped: ~/.claude/skills/<skill-name>/SKILL.md

Project skills are ideal for team conventions and should be checked into git.

Reload Behavior

• Skills are loaded from disk when Claude Code starts.
• If you create or edit skills while a session is running, restart Claude Code to pick up changes.

High-Value Claude Code Patterns

• Use skill-creator to bootstrap structure quickly.
• Keep deterministic logic in scripts; keep orchestration logic in SKILL.md.
• Use subagents in parallel to evaluate multiple skills against the same checklist.
• Treat skill reviews like code reviews (activation quality, edge cases, output contracts).

Pattern: Build + Review + Test Skills

A robust development loop in Claude Code:

1. Use an implementation skill (for example “adding CLI command”)
2. Use a review skill to check conventions and regressions
3. Use a test-generation skill to add coverage and run tests

This creates a predictable pipeline for feature work, not just one-off prompting.

Subagents + Skills in Claude Code

Subagents are useful when you want isolated context and parallel workstreams.

• Configure subagents with /agents (name, description, tools, model, prompt)
• Assign skills per subagent via the subagent skills field
• Subagents do not inherit parent conversation skills
• In subagents, configured skills are preloaded (full skill content is injected at startup)

This makes subagents ideal for dedicated reviewer/tester roles while the main agent focuses on implementation.

Tool Scoping for Subagents

Use least privilege:

• Reviewer subagent: read/search + optional bash
• Test runner subagent: read/search + edit/write + bash
• Keep destructive tools restricted unless explicitly needed

Constrained tool access improves safety and keeps delegated behavior focused.

Example Prompt Pattern

Use the skill-creator skill to evaluate these two skills against best practices:
1) generating-practice-questions
2) analyzing-time-series

Run evaluations in parallel subagents and produce a short scorecard:
- metadata quality
- workflow clarity
- edge-case handling
- output predictability

Common Pitfalls in Claude Code

• Installing a skill but forgetting to enable/reload it
• Over-trusting generated scaffolding without domain constraints
• Leaving script dependencies implicit
• Validating only happy paths
• Putting project-wide conventions only in skills instead of CLAUDE.md
• Assuming subagents inherit parent skills automatically