Auditing a deployed change. Marketing tweaked the system prompt last week — what exactly is different? Paste both versions and read the diff. A/B testing prompts. Two candidate prompts, one runs better on evals. Diff them to isolate what the difference might be doing. Reverting a regression. The latest prompt is worse than two iterations ago — which line did you change? Reviewing a teammate's edit. They sent you a "small tweak" to the system prompt — did they only touch the part they said they did? Migrating between model families. Adapting a prompt from GPT to Claude often means small wording changes — diff after rewriting to confirm the structure stayed the same.

This is a line-level diff. A single word changed in the middle of a long line marks the whole line as added+removed. For prose-level diff of a sentence, you may prefer a word-level tool. Trailing whitespace. Hidden spaces at the end of a line will mark it as different — useful or noisy depending on the case. Use the "Ignore trailing whitespace" toggle if you only care about visible content. Reordered blocks look like delete+add. If you moved a paragraph from position 1 to position 3, the diff shows it as removed at position 1 and added at position 3. There's no "moved" detection. Lines, not tokens. This diff doesn't speak token; it speaks lines. If your two prompts have the same content rewrapped to different line lengths, every line will look different. Normalise line breaks first. Privacy. Everything stays in your tab. Don't paste secrets into the tool's example placeholder, mind you — the placeholder text is hard-coded, not connected to your input.

Prompt Diff

Compare two prompts (or two versions of a prompt) side by side. Line-level diff with highlighted changes. Designed for iterating on LLM prompts.

Version A

Version B

View

Ignore trailing whitespace

Result

What is this for?

Prompt engineering is iteration: you write a prompt, test it, tweak one phrase, test again. After a dozen rounds, you've got a "version 1" and a "version 14" and no clean record of what changed where. This tool gives you that record on demand — paste any two prompts and see exactly which lines were added, removed, or left alone. No git, no setup, no upload.

When to use it

Auditing a deployed change. Marketing tweaked the system prompt last week — what exactly is different? Paste both versions and read the diff.
A/B testing prompts. Two candidate prompts, one runs better on evals. Diff them to isolate what the difference might be doing.
Reverting a regression. The latest prompt is worse than two iterations ago — which line did you change?
Reviewing a teammate's edit. They sent you a "small tweak" to the system prompt — did they only touch the part they said they did?
Migrating between model families. Adapting a prompt from GPT to Claude often means small wording changes — diff after rewriting to confirm the structure stayed the same.

Side-by-side vs unified

Side-by-side — A on the left, B on the right. Good when both versions are similar length and you want to scan visually.
Unified — single column with + / − markers, mirroring git diff output. Better for sharing the diff in a Slack message or for sparse changes.

Common gotchas

This is a line-level diff. A single word changed in the middle of a long line marks the whole line as added+removed. For prose-level diff of a sentence, you may prefer a word-level tool.
Trailing whitespace. Hidden spaces at the end of a line will mark it as different — useful or noisy depending on the case. Use the "Ignore trailing whitespace" toggle if you only care about visible content.
Reordered blocks look like delete+add. If you moved a paragraph from position 1 to position 3, the diff shows it as removed at position 1 and added at position 3. There's no "moved" detection.
Lines, not tokens. This diff doesn't speak token; it speaks lines. If your two prompts have the same content rewrapped to different line lengths, every line will look different. Normalise line breaks first.
Privacy. Everything stays in your tab. Don't paste secrets into the tool's example placeholder, mind you — the placeholder text is hard-coded, not connected to your input.