- Give Claude a check it can run, such as tests, a build or a screenshot comparison, so it can iterate without you as the only verifier.
- Explore and plan in plan mode first, then implement and commit. Skip the plan when you could describe the diff in one sentence.
- Be specific in prompts: name the files, point to existing patterns and describe a bug by its symptom.
- Keep CLAUDE.md short and limited to what Claude can't infer from the code. Use hooks for rules that must always run and skills for knowledge needed only sometimes.
- Manage context actively: clear between unrelated tasks, rewind bad turns, and send research to subagents.
- Scale up with headless `claude -p`, parallel sessions and worktrees, and have a fresh session review the work.
This is a summary in our own words, not the article itself. Read the original for the full text.