Claude Code best practices: CLAUDE.md, skills, hooks and plan mode

The Claude Code best practices that matter most are simple: keep a short project CLAUDE.md of facts every session needs, explore and plan before coding on anything non-trivial, clear the context between unrelated tasks, give the agent a check it can run, turn repeated workflows into skills, and enforce hard rules with hooks rather than polite instructions. Most of these come straight from Anthropic's documentation. The rest are what we have learned running our own company's work through coding agents.

Why do practices matter more than prompts?

Two developers with the same tool can get very different results. The difference is rarely a clever prompt. It is whether the agent has the right context, a way to check itself, and limits that stop it from doing damage. Those are set up once, in files your team commits, which is why they matter more than any one person's prompting style.

Eight practices that matter most

1. Keep CLAUDE.md short and factual

CLAUDE.md loads at the start of every session, so every line costs context on every task. The memory docs suggest adding to it when the agent makes the same mistake a second time, when a code review catches something it should have known, or when a new teammate would need the same context.

2. Explore, plan, then code

Anthropic's best practices recommend four phases: explore, plan, implement, commit. In plan mode the agent reads the code and proposes an approach without editing anything. You correct the plan, then let it implement.

The same page gives a useful rule for when to skip it: if you could describe the diff in one sentence, skip the plan. Planning is overhead that pays off on changes touching several files or an unfamiliar part of the codebase.

3. Manage the context window

The docs are direct about it: the context window holds the whole conversation, every file read and every command output, it fills fast, and performance degrades as it fills. In practice:

• Run /clear between unrelated tasks. A fresh start beats a long session carrying stale details.
• Restart after repeated corrections. If you have corrected the same thing twice, clear and start again with a better prompt that includes what you learned.
• Use subagents for investigation. Subagents run in their own context and report back a summary, which keeps the main session clean for the actual change.

4. Give it a check it can run

"Give Claude a check it can run: tests, a build, a screenshot to compare" is one of the most useful lines in Anthropic's guide. An agent that can run the tests sees its own mistakes and fixes them before you review. An agent without a check hands you work that looks right.

For teams this means the test and build commands belong in CLAUDE.md, and the permission rules should allow them without asking.

5. Turn repeated workflows into skills

A skill is a folder with a SKILL.md file: a short frontmatter block that says when to use it, then the instructions. The skills docs show that a skill at .claude/skills/deploy/SKILL.md creates a /deploy command for the project.

Good candidates for skills:

• Releasing a version (changelog, tag, deploy, smoke test).
• Adding a new API endpoint the way your codebase expects.
• Writing a migration with your naming and rollback rules.
• Reviewing a pull request against your team's checklist.

Commit project skills to the repository so every developer, and every agent session, follows the same steps.

6. Enforce hard rules with hooks

Instructions are requests. Hooks are enforcement. The hooks docs describe user-defined shell commands, HTTP endpoints and other handlers that run automatically at specific points in the agent's lifecycle. A PreToolUse hook runs before a tool call and can block it.

Typical team hooks:

Keep hooks fast and predictable. A slow hook on every tool call makes the whole team's experience worse.

7. Write specific prompts

The docs put it simply: the more precise your instructions, the fewer corrections you need. "Fix the login bug" is weak. "Users with a + in their email cannot log in, the check is in auth/validate.ts, add a test for that case first" gets a correct fix in one pass.

8. Use it outside the terminal too

• Non-interactive mode. claude -p "prompt" runs it in CI, pre-commit hooks or scripts.
• GitHub. With the GitHub Actions integration, mentioning @claude in a pull request or issue comment lets it analyze code, implement changes and push commits.

Start these after the interactive setup works well. Automation repeats whatever your setup gets wrong.

Frequently asked questions

Make the setup part of the codebase

The teams that get the most from coding agents treat the setup as code: reviewed, versioned and improved every time something goes wrong. Nactore sets up AI-native engineering for teams, from the first CLAUDE.md to skills, hooks and MCP servers. For the rollout side, read Claude Code for teams.

Want this set up for your codebase? Book a free 30-minute call.