Prompt Diff

About Prompt Diff

Paste two prompt versions side by side and get a line-by-line diff with inline highlights: red removals, green additions, blue modifications — Chinese marked per character, English per word. Line stats (unchanged / added / removed / modified) and an LCS similarity come along, clearly labelled as textual overlap rather than semantic similarity. The diff is our hand-rolled longest-common-subsequence implementation — zero third-party dependencies, running entirely in your browser.

FAQ

Why diff prompts instead of just reading version history?

Version history tells you WHICH version; a diff tells you exactly WHICH WORDS changed. Prompt engineering traps hide in small edits: swapping "concise" for "detailed", dropping a constraint line, changing the output format — each can swing model behaviour while being nearly invisible in two long texts. Locate every change with a diff first, then validate output shifts one by one; that's the standard iteration loop.

What does "similarity 85%" mean? Is higher better?

It's purely the longest-common-subsequence ratio: 100% identical, 0% disjoint, and anything between measures textual overlap. It says nothing about meaning — a full paraphrase keeps the sense while scoring maybe 50%. So no, higher is not "better": the goal of editing a prompt is editing it RIGHT, not editing it little. Use the number as a rough sense of change magnitude, read alongside the line diff.

Why not use an off-the-shelf diff library?

We avoid dependencies when we can (project discipline): the LCS core is ~60 lines and we tailor it for Chinese — no word boundaries, so inline highlights run per character for CJK and per word for Latin, while stock libraries default English-first. A hand-rolled pass also keeps the accounting (what counts as a modification vs a delete+add) fully transparent and testable. The trade-off: LCS is O(n·m), so don't diff entire books.