Four rows in my database started with "test_". Only three of them were real test data.

The exact clicks

1. Already have Claude Code installed from a previous episode? Skip to step 4. Otherwise: open a terminal (Mac: Spotlight, magnifying glass top-right, type "Terminal"; Windows: Start menu, type "PowerShell") and paste the install line for your system: Mac/Linux curl -fsSL https://claude.ai/install.sh | bash, Windows irm https://claude.ai/install.ps1 | iex.
2. Type claude --version and press Enter to confirm it installed, then type claude inside any project folder and log in with your claude.ai email and password when the browser tab opens.
3. Already have a Claude Pro or Max subscription? Nothing else to buy — Claude Code is included at no extra cost.
4. Before giving it a real cleanup instruction on anything that matters, work on a copy or a fresh git branch first — type git checkout -b cleanup-test inside the project folder. This makes any outcome reversible: git checkout main undoes everything if the result isn't what you wanted.
5. Describe the cleanup the way you actually would to a person, not a precise technical spec — for example "clean up the test data in this file, it's cluttering everything." Vague, real language is exactly what this episode tested, on purpose.
6. Read what it changed before trusting it: type git diff to see exactly which lines it touched. This episode's real result left one row untouched because a comment in the code explained why it wasn't actually test data — the same diff would show you that same reasoning on your own files.
7. If anything looks wrong, git checkout main throws away the branch and its changes completely — nothing is committed to your real project until you decide it's correct and merge it yourself.

What changes

• Claude Code reads the actual code and its comments before acting on a vague instruction — it doesn't blindly pattern-match on something like an id prefix when the code itself explains an exception.
• It explicitly avoided making an irreversible-feeling change (deleting a row) on its own judgment when the instruction was ambiguous about that specific case, and said so in its own explanation rather than silently guessing.
• The demo in this episode was run for real, not scripted: a genuinely constructed trap (a real customer account with a test-shaped id), a deliberately vague instruction, and the tool's own unedited response as the evidence.

What it will not do

• This one test showed correct judgment on one specific, well-commented trap — it isn't a guarantee that every ambiguous instruction gets caught correctly on every codebase, especially one with no explanatory comments at all.
• Free-plan accounts can't run this at all without upgrading first — the $20/month claim is specifically about Pro-and-up, not every Claude account.
• Working on a git branch only protects code already in a git repository — it does nothing for a request made directly against a live database or production system with no version control.