Back to Blog

DevGhost — Let AI Write Your Git Commit Messages Intelligently Using Gemini AI

May 15, 2024

We've all been there. You've just finished a solid chunk of work — refactored a service, fixed a gnarly bug, added a new feature — and then Git asks you for a commit message. And you type:

git commit -m "fix stuff"

Or worse:

git commit -m "asdfgh"

It's a universal developer sin. Commit history becomes a graveyard of meaningless messages that help nobody — not your teammates, not your future self, not anyone trying to bisect a bug three months later.

I built DevGhost to fix exactly this.

What is DevGhost?

DevGhost is a cross-platform CLI tool that automatically generates meaningful Git commit messages by analyzing your staged diff and sending it to Google's Gemini AI. It understands what you actually changed and writes a commit message that accurately describes it.

Install it once, use it everywhere:

pip install devghost

The Problem with Commit Messages

Good commit messages follow conventions like Conventional Commits — feat:, fix:, refactor:, chore: — with a clear, descriptive subject line. Writing these consistently takes discipline, especially under deadline pressure.

The result? Most codebases have commit history that looks like this:

- update
- fix
- wip
- fix again
- final fix
- FINAL fix for real
- okay this is the actual final fix

This is noise. It makes git log, git bisect, and code reviews significantly harder.

How DevGhost Works

The workflow is identical to your existing Git flow — just swap git commit for devghost commit:

# Stage your changes as usual
git add .

# Let AI write the commit message
devghost commit

DevGhost will:

  1. Run git diff --staged to capture everything you've staged
  2. Parse the diff into structured context (files changed, lines added/removed, code patterns)
  3. Send that context to Gemini AI with a carefully crafted prompt
  4. Stream the generated commit message back to your terminal
  5. Execute git commit -m "<generated message>" automatically

The whole thing takes under 3 seconds.

Example Output

Given a diff where you added JWT middleware to an Express app, DevGhost generates:

feat(auth): add JWT middleware for protected route authentication

- Implement verifyToken middleware using jsonwebtoken
- Apply middleware to /api/user and /api/dashboard routes
- Return 401 with descriptive error on invalid or expired tokens

That's a commit message that actually tells a story.

Prompt Engineering

The quality of the output depends entirely on how you frame the request to the model. After a lot of iteration, the prompt that works best is one that:

  • Gives the model the raw diff (not a summary — the actual lines)
  • Instructs it to follow Conventional Commits format
  • Tells it to be specific about what changed, not just that something changed
  • Caps the subject line at 72 characters
  • Uses imperative mood ("add", "fix", "remove" — not "added", "fixed", "removed")

The model I settled on is gemini-1.5-flash — fast, cheap, and accurate enough for this use case.

Secure API Key Storage

DevGhost stores your Gemini API key securely using the OS keychain (via the keyring library) — not in a .env file or plain config that could accidentally get committed.

First-time setup:

devghost setup
# Enter your Gemini API key when prompted
# It's stored in your system keychain, never on disk as plain text

After that, it just works.

Installation & Setup

Requirements: Python 3.8+

# Install from PyPI
pip install devghost

# One-time setup — stores your API key securely
devghost setup

# Use it
git add .
devghost commit

Get your free Gemini API key at aistudio.google.com.

What I Learned Building This

Diff parsing is harder than it looks. Raw git diff output has a lot of noise — chunk headers, file metadata, binary file indicators. Stripping this down to just the meaningful code changes, without losing context, required careful regex work.

Prompt length matters. Large diffs can easily exceed token limits. DevGhost truncates intelligently — it keeps the first and last N lines of each file's diff rather than cutting at a hard character limit, so the model sees the beginning and end of each change.

Streaming makes it feel fast. Even if the API call takes 2 seconds, streaming the tokens as they arrive makes the UX feel instant. Python's google-generativeai SDK supports streaming with a simple stream=True flag.

What's Next

A few things on the roadmap:

  • devghost amend — regenerate a message for your last commit
  • Support for OpenAI and Claude models as alternatives to Gemini
  • A --dry-run flag to preview without committing
  • VS Code extension so you never have to leave the editor

Try It

pip install devghost

Source code on GitHub. Stars appreciated 🙂

If you have feedback or want to contribute, open an issue or reach out on LinkedIn.

GitHub
LinkedIn