ArrayKit

Prompt Diff

Compare two versions of a prompt word by word, with the token delta the edit costs.

Prompt Diff

Compare two versions of a prompt word by word, with the token delta the edit costs.

Both versions are compared in the browser. Prompts you are still iterating on never leave the machine.

Lint the new version

About Prompt Diff

A line diff is the wrong tool for a prompt. Prompts are prose: rewrap a paragraph and every line changes while the meaning is untouched, so the diff is all noise. This compares by word instead, which is what you actually edited, and reports the one number a code diff never shows — how many tokens the change adds or removes from every single request from now on. Long unchanged stretches collapse to a marker so the edits are readable, and the whole thing can be exported in a unified style for a review.

Features

  • Word-level comparison, so rewrapping is not a change
  • The token delta per request, which is what the edit really costs
  • Added, removed and unchanged word counts
  • A similarity percentage for how much survived
  • Long unchanged runs collapsed, with context kept either side
  • Both sides reconstructible exactly from the diff
  • A guard against pasting something too large to diff quadratically
  • Exports in a unified style for a code review

How to use the Prompt Diff

  1. Paste the previous version on the left
  2. Paste the new version on the right
  3. Read the highlighted changes and the token delta
  4. Export the unified diff if you need it in a review

Example

Input

Be concise.

Output

Keep replies under four sentences.

A four-word change that alters the behaviour and the cost.

Common errors & troubleshooting

  • It says the inputs are too large. — Word-level diffing builds a table proportional to both lengths. Compare a section at a time rather than two whole documents.
  • The whole thing shows as changed. — Usually a difference in invisible characters — smart quotes, non-breaking spaces, or a different line ending pasted from a document.
  • The token delta looks approximate. — It is. It is a character-based estimate, close enough to tell a five-token edit from a five-hundred-token one, which is the decision it exists to support.
  • Collapsing hides something I want to see. — Turn it off. Everything is still there; collapsing only replaces long unchanged runs with a marker.

Frequently asked questions

Why not use an ordinary text diff?
Because prompts are paragraphs, not lines. Reformatting one sentence rewraps the paragraph and a line diff reports the whole thing as changed, which buries the actual edit.
Why does the token delta matter?
A system prompt is sent with every request. Twenty tokens added is twenty tokens on every call for the life of the prompt, which is a different kind of decision from twenty tokens in a document.
Can I use it for anything other than prompts?
It works on any prose — tool descriptions, skill instructions, project memory. Anywhere a line diff would be dominated by rewrapping is a good fit.
Does the diff preserve my text exactly?
Yes. Reassembling the unchanged and removed pieces reproduces the original exactly, and the unchanged and added pieces reproduce the new version, whitespace included.
What does the similarity percentage mean?
The share of the longer version's words that survived unchanged. It is a quick way to tell a wording tweak from a rewrite, particularly when reviewing someone else's edit.

Related tools

More AI & Agents tools

See all AI & Agents tools · All tools