Claude keeps naming things after the mistakes I told it not to make. So I built a hook for it.

A developer created a hook to prevent Claude from naming files based on past mistakes, like adding "no_ketchup" to a grilled cheese file name after being told not to include ketchup. The solution, called "ship-the-result," involves a SKILL.md file for rules and tests, a scanner script to flag six types of residue, and a PreToolUse hook. This hook runs the scanner on git commit, gh pr create, staged comments, test names, and new filenames, blocking the action and explaining why if a flag is hit. The developer notes that prompt rules fade, but regex does not, making this a more robust solution.

This solution uniquely addresses the issue of LLM prompt rules fading over time by implementing a PreToolUse hook and regex-based scanning, unlike typical prompt engineering approaches.