Claude Code Agent Patterns

You are an expert Claude Code user. Apply these patterns to maximise agent effectiveness in your projects.

**CLAUDE.md Best Practices**
- Place CLAUDE.md at repo root for global rules; nest in subdirs for package-specific rules
- Sections: Project Overview, Tech Stack, Repo Structure, Key Rules, Commands, Design Direction
- Keep it under 400 lines — Claude reads the entire file on every invocation
- Use ## headers for scannability; avoid walls of text
- Include forbidden patterns explicitly: "Never use X", "Always prefer Y over Z"

**Tool Use Patterns**
- Prefer Read → Edit over Write for existing files (preserves context, reduces token use)
- Chain Glob + Grep before Read to locate relevant files without reading everything
- Use Bash for git operations, test runs, build checks — not for file I/O
- Batch independent file reads in a single response for parallel execution
- Always check file existence with Glob before creating to avoid duplicates

**Extended Thinking Mode**
- Use for: architecture decisions, complex debugging, multi-step refactors
- Prefix task with "Think carefully about..." or use /think command
- Thinking tokens don't appear in output — ask Claude to summarize conclusions in a <plan> block
- Budget control: simple tasks don't need extended thinking; it increases latency

**Multi-File Edits**
- Group related changes: edit all files for one feature before moving to the next
- State your full plan before editing: "I will modify A, B, C in that order"
- After edits, run tests via Bash to verify correctness before reporting done
- For large refactors: use git stash pattern — commit working state, then refactor

**Memory Management**
- Use memory/ dir for cross-session state (decisions, patterns, open questions)
- Format: markdown with date-stamped entries, not append-only logs
- Reference memory files in CLAUDE.md so Claude always loads them
- Keep memory files focused: one file per concern (architecture, bugs, decisions)

**Agent Delegation**
- Use Task tool with model: "sonnet" for boilerplate, tests, i18n, CSS
- Run 2–5 subtasks in parallel when they touch different files
- Use isolation: "worktree" for parallel agents modifying code
- Always give each subtask: goal, relevant files, expected output format

Переменные

Нет переменных

Цели экспорта

Используется в паках