Prompt diff: compare two prompt versions

Prompts change all the time: a rule is added after a bad answer, an example is swapped, a sentence is reworded. Small edits change how a model behaves, so it helps to see exactly what changed between the version that worked and the one you are about to ship. A general text diff shows the words. This tool also shows the two things a general diff misses: what the edit does to your token bill, and whether it broke the template your code fills in.

How the comparison works

The diff uses the Myers algorithm, the same approach behind git diff, through the open-source jsdiff library. In word mode each prompt is split into words (letters in any script, digits and underscores), punctuation marks and the spacing between them, so a changed number or a swapped comma shows up on its own. In line mode each line is one unit, which suits rule lists and few-shot examples. “Ignore whitespace” compares words without the spaces and line breaks around them (and lines with spacing collapsed), and “Ignore case” compares in lower case; either way, each side is still shown exactly as you wrote it. If two very long versions are too different to compare word by word quickly, the tool falls back to whole lines rather than freezing the page.

Tokens and cost

Both versions are counted with the tokenizer of the model you choose, in a background worker so typing stays smooth. Counts are exact for OpenAI models and for open-weight models whose tokenizer is public, such as DeepSeek, Qwen and Mistral. Anthropic and Google don’t publish their current tokenizers, so for Claude and Gemini the numbers are estimates made with OpenAI’s o200k_base and marked with ≈; the size of the change is still a fair guide. The cost is the prompt billed as input at the model’s standard price, refreshed daily from public price lists, including the higher long-context price when a prompt passes the model’s threshold. The reply, caching and batch discounts are left out because the wording of the prompt doesn’t decide them; the cost calculator covers the whole request.

A longer prompt costs more on every call, and an edit near the start of a prompt also resets prompt caching, because providers only reuse an identical prefix. Put stable instructions first and the parts you change often last.

Template variables

Most prompts in production are templates. The tool finds placeholders in two styles: double braces such as {{customer_name}} (Mustache, Handlebars, Jinja and most prompt tools, including filters such as {{ name | upper }}) and single braces such as {language} (Python’s str.format, f-strings and LangChain, including {user.name}, {items[0]} and format specs like {price:.2f}). It lists variables added, removed and kept, and those used a different number of times. A renamed variable shows up as one removed and one added, which is the case that breaks things: depending on the engine, a missing value raises an error (Python), renders as an empty string (Mustache, and Jinja by default) or reaches the model as the literal placeholder. JSON objects, code with spaces inside its braces, ${…} interpolation in JavaScript or shell, LaTeX commands and Handlebars helpers such as {{#each}} aren’t counted, though a bare {x} in code looks the same as a Python placeholder and is listed. In Python templates, doubled braces are an escaped literal brace, not a variable; the tool can’t know which engine you use, so it lists both styles separately.

What it doesn’t do

It compares wording, not behaviour. Two prompts that differ by one word can produce very different answers, and a big rewrite can change nothing that matters. To know whether an edit helped, run both versions on the same set of test inputs and compare the outputs. Your prompts stay in your browser: they aren’t uploaded, stored or added to the page address, which is also why there is no share link.