Claude Code's safety filter keeps blocking a harmless hobby project
A developer reports repeated safety-classifier interruptions in Claude Code while building a hobby app that has nothing to do with . The 's rewind feature reportedly fails about 90% of the time, so the developer switched to the CLI (claude -c) to roll back sessions instead. Even after manually checking the chat for flagged words like "bypass" or "hack" and finding none, the block recurred mid-session, after several files had already been written and multiple tool calls had already been approved.
As a result, the developer canceled their Claude MAX and is considering switching to a different model, such as Kimi K2, once the already- period ends in August. They note that (running on GPT-5.1), used in parallel to cross-check the same work, has never triggered a similar block.
Key points
- The 's rewind feature reportedly fails to work about 90% of the time
- Safety blocks can recur mid-session even after checking for and finding no flagged words
- The developer canceled Claude MAX and is weighing a switch to another model like Kimi K2 in August
- (GPT-5.1), used for cross-checking the same work, never triggered a similar block