Claude Code habits that actually stick

The shift

I've been using Claude Code daily since March 2026. Upgraded to Max in April. After using it seriously — not just for autocomplete, but for architecture, debugging, content, the whole thing — here are the habits that actually stuck.

Six habits that matter

1. Plan before you touch code.

Any task that touches more than one file, or where the scope is unclear, starts with /plan. The plan agent explores the codebase, designs the approach, and writes a step-by-step implementation. This is not optional overhead — it's the difference between a two-hour task and a two-day mess.

2. Write CLAUDE.md rules the moment you discover them.

The second you notice Claude doing something wrong repeatedly — wrong pattern, wrong assumption, wrong tool — stop and write a rule. Don't batch them up for later. Later never comes, and meanwhile you're correcting the same mistake ten times.

3. Skills for repeatable behavior, not chat instructions.

If you want Claude to do something consistently across sessions, encode it as a skill file. Chat instructions evaporate. A skill file in ~/.claude/skills/ does not.

4. Name every session.

Every session gets a snake_case title at the start. refactor_auth_flow, debug_supabase_rls, build_checkout_page. This makes memory search actually work. Without names, sessions blur together and context is lost.

5. Match complexity to delegation level.

The mental model: one file is inline work, two to five files goes to a Sonnet agent via the Agent tool, cross-domain or architectural work goes to the orchestrator. Using the right level saves tokens and gets better results. Don't pull in the orchestrator for a one-file fix.

6. Call the advisor sparingly.

The advisor (Opus) is expensive. Only worth it for architecture decisions, security-sensitive design, or when you've tried twice and failed. Don't use it to rubber-stamp obvious decisions — that's just burning money for comfort.

What doesn't work

Vague instructions. "Be careful with the database." "Always test your changes." Claude needs specific and actionable. If it matters, write the rule. If you can't write a specific rule for it, it's not a rule — it's a vibe, and vibes don't ship.