PromptFixer lint · score · rewrite

Desktop app for Windows and macOS · Open source, MIT

Lint, score and rewrite your LLM prompts. Offline.

PromptFixer scores a prompt against about thirty prompt-engineering rules, then a built-in local model rewrites it and the rewrite is scored again, so you can see whether it really got better. It works offline, nothing you type leaves your machine, and an automated test checks that on every push.

Download for Windows View on GitHub

Windows 10 or 11, 64-bit. The 2.5 GB model downloads on first run; after that it works offline.

The PromptFixer window: a prompt scored 42 in the editor and its rewrite scored 93 on the Fixed tab, with one passage marked to keep and one to change.
A five-word prompt scored 42; its rewrite scored 93. One passage is marked to keep and one to change, ready for Fix again with my marks.

How it works

Three steps, and the score stays comparable through all of them.

  1. Paste

    Paste a prompt and it is scored instantly against about thirty prompt-engineering rules, with the findings grouped by category and severity. The linter is deterministic and local: no model is involved, and nothing is sent anywhere.

  2. Fix

    Press Fix prompt and the local model rewrites it. The rewrite is scored again with the same rules, next to a word-level diff and a change log that explains every edit. Choose a strength, Light touch, Balanced or Full rebuild, and a guard retries a rewrite that changed too much or pasted its own instructions into the output.

  3. Mark and fix again

    Not quite what you wanted? Select a passage of the rewrite and mark it Keep it — I loved it or Change it — I didn't like it, then press Fix again with my marks. What you kept comes back word for word, what you disliked is reworded, and the result is still scored against the prompt you wrote.

What you get

Scoring is instant and local; the rewrite needs the model downloaded once.

  • Private by design

    The default model runs on your machine. There is no telemetry, no analytics, no crash reporting, no licence check and no update check. docs/PRIVACY.md names every host the app can reach, and when.

  • Built-in local model

    Qwen3-4B-Instruct-2507 (2.5 GB) by default, with a 1 GB lite tier for machines with 4 GB of RAM and a 4.7 GB quality tier. CUDA, Vulkan or Metal is picked automatically, with a CPU fallback.

  • Watch the rewrite arrive

    The rewrite streams into the Fixed tab as the model writes it, and the spinner names which wait you are in: queued behind another job, loading the model, writing, or a second attempt after a failed quality check.

  • A library that teaches the model

    Save a prompt with its rewrite, and your saved rewrites are shown to the model as examples of how much change you like. Export the library as JSON and import it on another machine, merged without overwriting anything.

  • Diff and change log

    Every rewrite comes with a word-level diff and a change log explaining each edit, and is scored before and after with identical rules, which is what makes the comparison meaningful.

  • Cloud providers, optional

    Anthropic, OpenAI, OpenRouter, Gemini, Ollama or any OpenAI-compatible endpoint, switched on with a .env file. Off by default, and nothing requires them.

  • Lint prompts in CI

    npm run lint:prompts runs the same rules over your prompt files and exits non-zero when one falls below the bar. No server, no model, no API key.

  • Tested on every push

    Unit and component tests, an acceptance suite that drives the real app in a browser and in Electron, and a visual check with Elastishot that names the element behind every difference it finds.

Download

One file, no account, nothing to install first. Windows 10 or 11 (64-bit), or macOS 13 or later.

Latest release

Published

Download the latest release from GitHub

File
Size
SHA-256
SHA256SUMS.txt

Full install guide All releases

Before you run it

Check the download

In PowerShell, from the folder you saved it to. The result must equal the SHA-256 of the release, shown on the download card, which is also in SHA256SUMS.txt on its release page. 

(Get-FileHash .\PromptFixer-<version>-win-x64.exe).Hash.ToLower()

On a Mac, in Terminal; the other image differs only in its file name.

shasum -a 256 ~/Downloads/PromptFixer-<version>-mac-arm64.dmg

Not code-signed yet

Windows shows a blue "Windows protected your PC" screen because the installer is not code-signed. Click More info, then Run anyway, once the hash matches.

macOS refuses the first launch with "Apple could not verify PromptFixer is free of malware": the app is ad-hoc signed, not notarized. Click Done, open System Settings, Privacy & Security, and press Open Anyway. Once.

Upgrading

Your library and the downloaded model live outside the app and survive uninstalling and reinstalling.

What's new

Every release, newest first, with the notes as they were published.

The release notes could not be loaded on this page. Read them on the GitHub releases page.

Full changelog on GitHub