tldr

Search tldr

By Anthropic in AI2 min read

Best practices for Claude Code - Claude Code Docs

Anthropic’s guide to operating Claude Code effectively: constrain context, give the agent verifiable goals, and use planning, tooling, and independent review to scale reliable work. It emphasizes that autonomy works best when paired with explicit checks and disciplined session management.

Share

Best practices for Claude Code - Claude Code Docs

Author: Anthropic | Published: Unknown | Generated: 2025-04-21 | Domain: anthropic.com Tags: ‘#claudecode’ ‘#aiagents’ ‘#developer-tools’ ‘#context-management’ ‘#automation’

TLDR

Claude Code is most effective when treated as an autonomous engineering agent with explicit verification, well-scoped context, and clear operational boundaries. Anthropic recommends giving it pass/fail checks such as tests, builds, linters, or visual comparisons; separating exploration, planning, and implementation for complex changes; and aggressively clearing or delegating context-heavy research. Teams can scale this approach through permissions, hooks, skills, MCP servers, parallel worktrees, non-interactive runs, and fresh-context adversarial reviews.

Key Takeaways

  • Verification enables autonomy: Give Claude tests, build commands, linters, fixture diffs, or screenshot comparisons so it can iterate to a concrete pass/fail result; Stop hooks can block completion until a deterministic check passes, though Claude Code ends a turn after 8 consecutive blocks.
  • Manage the context window as a primary resource: Conversations, file reads, and command output all consume context, and quality can degrade as it fills. Use /clear between unrelated tasks, /compact for targeted summarization, /btw for disposable side questions, and subagents for investigation.
  • Plan before multi-file or uncertain work: Use plan mode to explore without edits, create and review a detailed plan, then implement and verify it. Skip planning for small, obvious changes whose diff can be described in one sentence.
  • Persist only high-value instructions: Keep CLAUDE.md concise and focused on non-obvious project-wide commands, conventions, workflows, architecture, and gotchas; use on-demand skills for specialized knowledge to avoid bloating every session.
  • Scale with isolation and independent review: Use worktrees, parallel sessions, /batch, or claude -p for fan-out work, restrict unattended runs with --allowedTools, and have a fresh subagent review diffs for requirement or correctness gaps rather than style preferences.

Images & Media

None present in article.

  • How Claude Code works — Explains the agentic loop, tools, and context management model.
  • Plan mode — Documentation for exploring and planning without making edits.
  • CLAUDE.md files — Guidance on persistent project instructions and import rules.
  • MCP servers — Connects Claude Code to external tools and services such as Notion, Figma, and databases.
  • Non-interactive mode — Runs Claude with claude -p in CI, scripts, and other automated workflows.
  • Checkpointing — Details restoring conversation and code state with /rewind.

Keep reading

GitHub - Andyyyy64/whichllm: Find the local LLM that actually runs and performs best on your hardware. Ranked by real, recency-aware benchmarks, not parameter count. One command, run it instantly. · GitHub

whichllm is a Python CLI that detects or simulates hardware, then recommends runnable local LLMs using benchmark quality, recency, memory fit, and estimated speed—not parameter count alone. It also supports hardware planning, upgrade comparisons, instant model chat, and scriptable JSON output.