Claude Code configuration notes for CCAR-F

CLAUDE.md hierarchy

• User-level (~/.claude/CLAUDE.md) is personal—teammates will not get it via git.
• Project-level (CLAUDE.md / .claude/CLAUDE.md) is the shared source of truth.
• Use @import and .claude/rules/ to keep standards modular; path globs load rules only when relevant files are touched.
• /memory helps verify what is actually loaded when behavior looks inconsistent.

Commands & skills

• Project commands live in .claude/commands/; personal ones in ~/.claude/commands/.
• Skills (.claude/skills/*/SKILL.md) support frontmatter like context: fork, allowed-tools, and argument-hint.
• Forked skill context keeps verbose exploration from polluting the main thread.

Plan mode vs direct execution

• Plan mode for large/architectural/multi-approach work.
• Direct execution for narrow, well-scoped fixes.
• Explore subagents for noisy discovery so the main context stays clean.

Hooks patterns that matter

• PreToolUse / interception hooks: gate policy-sensitive calls before they run.
• PostToolUse: normalize tool results (dates, status codes) before the model reasons on them.
• PreCompact: archive transcripts before compaction discards detail.
• Prefer dedicated lifecycle hooks over inventing fake tool names for non-tool operations.

MCP config reminders

• Shared servers → project .mcp.json with ${ENV} expansion for secrets.
• Personal experiments → user-level Claude config so they do not ship to the team.