A Claude Code hook that swaps out the AI's most grating pet phrases

Developer Johanna Larsson has grown tired of Claude’s verbal tics — the model’s habit of calling everything an “honest take,” a “load-bearing seam,” or telling users “you’re absolutely right.” Rather than just complain, she wrote a short fix that intercepts and rewrites the offending words on their way to the screen.

The solution leans on Claude Code’s MessageDisplay hook, which fires on outgoing text. A small Python script reads the streamed delta, runs a set of regex replacements against a dictionary of phrases, and returns the doctored content. Her example dictionary is deliberately absurd — “load-bearing” becomes “cooked,” “honest take” becomes “spicy doodad,” and “you’re absolutely right” is rewritten as “I’m a complete clown.” Setup is three steps: drop the script in ~/.claude/hooks, mark it executable, and register it in settings.json under the hooks block. Hooks load at startup, so a new session picks up the changes.

The piece is played for laughs, but it doubles as a practical demonstration of the hook system’s reach. Because MessageDisplay operates on rendered output rather than the model itself, users can enforce their own vocabulary preferences client-side without touching prompts or model behavior — and the same mechanism could just as easily be pointed at more useful transformations than joke substitutions.