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.
This matters more than it sounds. Prompt engineering is essentially a debugging loop: you change something, observe the output, decide whether it improved, and iterate. Without a reliable diff, you end up comparing your mental model of the prompt against reality — a surprisingly unreliable process, especially after many small edits.
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.
The key advantage of word-level diffing over line-level (the default in most code diff tools) is that prompts are prose, not code. A single sentence often changes by one or two words, and a line diff would flag the entire sentence as changed — hiding the actual edit. Word-level diffing shows you exactly which token moved.
When to use this tool
- After a prompt edit produces an unexpected change in model output, to confirm exactly what you changed
- When collaborating with a team on a shared prompt and you receive a new version with no changelog
- When rolling back a prompt that regressed, to confirm the rollback landed correctly
- Before and after compression or reformatting, to verify no meaningful instructions were dropped
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.
- Correlate diffs with behaviour notes. The diff tells you what changed; your notes tell you what the output did. Together they build up a causal map of your prompt’s behaviour over time.