5. Agent Skills — Claude Code Configuration¶

Domain 3 (20%) — Claude Code Configuration & Workflows

CLAUDE.md configuration hierarchy — Three levels — which one takes precedence where

~/.claude/CLAUDE.md          ← User-level (personal, NOT version controlled)
.claude/CLAUDE.md            ← Project-level (shared via git ✅)
  or root CLAUDE.md
subdirectory/CLAUDE.md       ← Directory-level (applies to that folder only)

Critical Rule

User-level settings are INVISIBLE to teammates. Always put team config at project level.

Use @import syntax to keep CLAUDE.md modular:

@import ./standards/api-conventions.md
@import ./standards/testing.md

.claude/rules/ — path-scoped rules — Best answer for "apply different conventions to different code areas"

---
paths:
  - "src/api/**/*"
---
# API conventions apply here
Always use async/await with structured error handling...

YAML frontmatter with paths glob patterns
Rules only load when Claude is working on matching files
Best answer for "apply different conventions to different code areas"

Commands & Skills — location and scope — Project vs personal, version controlled vs not

Location	Scope	Version Controlled?
`.claude/commands/`	Project (all teammates)	✅ Yes
`~/.claude/commands/`	Personal only	❌ No
`.claude/skills/`	Project skills	✅ Yes
`~/.claude/skills/`	Personal skill variants	❌ No

SKILL.md key frontmatter options:

context: fork — most important for exam. Runs skill in isolated subagent — output doesn't pollute main conversation.
allowed-tools — restricts what tools the skill can call (least privilege)
argument-hint — UX hint shown to user

Plan mode vs direct execution — When to think before acting

Use Plan Mode	Use Direct Execution
Large-scale changes (45+ files)	Single-file bug fix with clear stack trace
Multiple valid implementation approaches	Well-understood, clear-scope change
Architectural decisions (microservices split)	Adding a simple conditional
Library migrations with unknown deps	Known refactor with clear target

Pattern

Plan mode to explore → Direct execution to implement

CI/CD integration + iterative techniques — The -p flag and best practices for automation

-p / --print flag = the ONLY correct way to run Claude Code in CI (non-interactive, prevents hanging)
--output-format json + --json-schema = machine-parseable findings for PR comments
Session isolation: fresh instance reviews code better than the one that wrote it
CLAUDE.md provides project context (testing standards, fixtures) to CI-invoked Claude
/memory command: shows which memory files are currently loaded — use to diagnose inconsistent behavior across sessions

Iterative refinement techniques:

Concrete examples beat prose: 2–3 input/output pairs > any description
Test-driven: Write tests first, share failures to guide fixes
Interview pattern: Have Claude ask questions BEFORE implementing in unfamiliar domains
Interacting issues: Fix in one message. Independent issues: Fix sequentially.