Which single line in a CLAUDE.md file makes the biggest difference?
People keep tuning their CLAUDE.md files. If you had to pick one instruction that changed how Claude Code behaves the most, which one would it be, and why does it work?
The lines that pay off most are the ones that change a habit, not the ones that describe the project. A rule like "run the tests and show the output before saying something is done" stops the most common failure: the agent claiming success it never checked.
Second best is anything the model cannot guess from the code: the exact command to build and test, folders it must not touch, and decisions already made. These save a wrong attempt on every new session.
Keep the file short. CLAUDE.md is loaded into every session, so each line costs context every time. If a rule only matters for one kind of task, it fits better in a skill that loads when that task comes up.
When a rule keeps getting ignored, make it concrete and checkable. "Be careful" does nothing; "never edit files under migrations/" is something the model can follow and you can verify.
Listings mentioned
- Verification before completion · skill by obraTurns "check before you claim it's done" into a full skill with a clear rule.
- Context engineering · skill by addyosmaniCovers what belongs in a rules file like CLAUDE.md and what should load only per task.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.