[Bug] File write operations skip read-before-write check and fail silently

Bug Description
You're paying for a tool that markets itself as an autonomous coding agent, and then you have to write a manual telling it how to use its own tools correctly. That's like buying a car and having to tape a note to the steering wheel that says "TURN THE WHEELS BEFORE DRIVING INTO A WALL." The read-before-write thing isn't some edge case preference — it's the most basic file operation pattern in computing. Every text editor written since the 1970s knows to check file state before writing. The fact that Claude Code needs a .claude.md rule to not skip this step is a product deficiency, not a configuration opportunity. And the silent failure part is worse. If it fails, it should scream at you. Instead it just quietly eats tokens, does some insane delete-and-recreate workaround, and you don't find out until you're looking at a mangled file 20 minutes later. That's not AI being dumb — that's bad error handling in the tooling layer that Anthropic ships around the model. The .claude.md rules file is a workaround for broken defaults. It shouldn't need to exist for basic operational hygiene. The defaults should be: read before you write, fail loudly, verify your work. That those aren't baked in from day one is an engineering prioritization problem on Anthropic's side. The Fortune article from literally today suggests the pressure is mounting — enough high-profile users are complaining publicly that it's becoming a PR issue. Hopefully that accelerates the fixes. But you shouldn't have to write a behavior contract for a tool you're paying for just to get it to not destroy your files.

Environment Info

• Platform: win32
• Terminal: pycharm
• Version: 2.1.108
• Feedback ID: 26ff6283-7935-48c4-9e91-c03d24e83a19

Errors

[]