Claude Code best practices: what to set up, and what to leave alone

Anthropic's best practices page (read 6 November 2026) puts three habits first, in our summary: First, tell Claude what it cannot see in your code, such as the exact test commands, a deploy process or a decision made outside the repository. Second, give it a check that comes back pass or fail, so it can fix its own mistakes instead of you finding them. Third, for rules that cannot have exceptions, use settings that Claude Code enforces, such as hooks and deny rules, rather than a note in CLAUDE.md that Claude is allowed to weigh. Anthropic also says its defaults already work well for a first plain prompt, so add setup only where your project needs it.

Name the file, describe the symptom and say what "fixed" looks like. Anthropic's page sets "fix the login bug" against a prompt that says users report login fails after the session times out, points at token refresh in src/auth/, asks for a failing test first, then the fix. It also says that when you already have the thing Claude needs, such as a file, a screenshot or command output, you should hand it over instead of describing it. On scope, the page says Claude tends to read a request conservatively, so saying "be ambitious" and listing what you want, such as empty and error states, tells it how much you hoped for.

Anthropic's page splits instructions into two kinds. A prompt, CLAUDE.md or a skill is read by the model, which usually follows it but is not guaranteed to. A hook, a deny rule or a managed setting is applied by Claude Code every time, whatever the model decided. The page's advice is that if something has to happen every time, make it a hook, a deny rule or a managed setting, and keep CLAUDE.md for guidelines. Its examples of rules worth moving are never reading .env, always running the formatter after an edit, and not touching the migrations folder. It adds that deny rules hold in every permission mode.