Prompt version differ
When you iterate on a prompt and the model’s behaviour shifts, the first question is always “what did I actually change?” This tool runs a word-level diff between two prompt versions and highlights every addition and removal, so you can tie a behaviour change to a specific edit instead of guessing.
How it works
Both versions are tokenised into words and whitespace. A longest-common- subsequence alignment finds the shared backbone between them; anything present only in version A is marked removed (struck through, red) and anything only in version B is marked added (highlighted, green). Unchanged tokens render neutral. A small summary counts how many words were added versus removed so you can gauge the size of the change at a glance.
Tips and notes
- Diff one change at a time. If you tweak the role, the format instructions, and the examples all at once, you cannot attribute the result to any single edit — version your prompts after each meaningful change.
- Keep an external history. Pair this with a notes file or version control; the differ compares two versions but does not store them for you.
- Watch whitespace edits. Added line breaks and reformatting show up as changes here, which is useful when formatting affects model parsing.
- Use it in reverse too. Swapping A and B turns additions into removals, handy for understanding a rollback.