Show Dev: Prompt Diff — Compare LLM Prompts by Token Count & Cost

#showdev #llm #tools #ai

title: "Show Dev: Prompt Diff — Compare LLM Prompts by Token Count & Cost"

tags: ["showdev", "llm", "tools", "ai"]

I built Prompt Diff — a free, offline-first tool for comparing two LLM prompts side-by-side, showing token delta, cost difference, and a text diff view. No API key, no data leaves your browser.

Live tool: https://citriac.github.io/prompt-diff.html

Why I built it

When I'm iterating on system prompts or user messages, I often want to know:

Did my revision actually reduce token count?
How much more expensive will 1,000 calls be with the new prompt?
What exactly changed between v1 and v2?

Copy-pasting into a token counter one at a time and doing mental math isn't fun. Prompt Diff does it all in one view.

What it does

Token & cost delta — Instantly shows how many tokens changed (+/-) and the estimated cost difference per single call and per 1,000 calls. Supports 19 models across OpenAI, Anthropic, Google, Meta, DeepSeek, Mistral, and Cohere.

Three diff modes:

Word diff — highlights individual added/removed words inline (like GitHub)
Line diff — classic +/- line view for structured prompts
Char diff — character-level for catching subtle rewording

The delta banner — four big numbers at a glance: token delta, cost delta, % change, char delta. Color-coded red (increase) / green (decrease).

Copy diff report — generates a Markdown table you can paste into a PR description, Notion doc, or Slack.

Share via URL — encodes both prompts in the URL for short prompts, so you can send a diff link to a teammate.

Export as Markdown — full export with stats table + both prompts.

Technical bits

The diff engine is a pure-JS LCS (Longest Common Subsequence) implementation — no libraries. Works at word, line, or character granularity. Token counting uses a character-ratio heuristic tuned per model family (GPT-4: ~3.5 chars/token, Gemini: ~4.0, etc.) — same approach as my Token Counter.

Everything runs in the browser. No server. No cookies. Bookmark it, use it offline.