skip to content

Search

Introducing kommit: Generate Meaningful Git Messages with AI

4 min read

AI-powered CLI tool for generating conventional commit messages. Solves the struggle of writing meaningful commits during rapid development.

# ai
# cli
# git
# productivity
# tools
# workflow
Not in series

Several months ago, I found myself needing to ship code really fast. I was moving so fast, even writing readable commit messages felt like a luxury. Lots of breaking changes were required with a tight deadline.

The Problem

From my past experience, I found it particularly difficult to make good commits when I was dealing with:

  • Early-stage code → because I’m unsure of the final implementation.
  • Large changes → because describing or summarising large changes succinctly is difficult.
  • Rapid prototyping or MVP → because code evolves faster than your thoughts (or the requirements change faster than you can code 🥁).

In these situations, there is a mental conflict in my head saying, I know exactly what I changed 😏, but why is it so hard to describe it? 🤔 I take commit messages seriously as a courtesy for others (or future me) that may review my code so I can run into analysis paralysis in such situations.

That’s when I thought: if LLMs can write code, why not let them handle my commits too?

The Solution

Since LLMs are so good at generating code I thought a good commit message can be generated just from diffs. My solution is kommit. When I started building this tool, which was early March, I wasn’t aware of Claude Code, or any other similar AI-powered CLI tools.

Some good feedback and feature requests from colleagues have led to a few releases so far, currently sitting at v0.3.1.

I’ll guide you through how to install, set up and run it. If you’d like to explore at your own pace, README covers most of what’s discussed below.

Install

Installing is straightforward:

  • Using homebrew: 
    brew install cowboy-bebug/tap/kommit
  • From source: 
    git clone https://github.com/cowboy-bebug/kommit.git
cd kommit
make install

Setup

To set up, either of these environment variables should be exported with the KOMMIT_ prefixed one taking precedence:

export OPENAI_API_KEY="sk-..."
export KOMMIT_OPENAI_API_KEY="sk-..."

Then run:

git kommit init

which creates a config file .kommitrc.yaml that looks like:

llm:
  model: gpt-4o-mini
commit:
  types:
    - feat
    - fix
    - docs
    # ... other conventional commit types
  scopes:
    - ui
    - api
    - auth
    # ... project-specific scopes

Usage

First, files or diff chunks must be staged. Then simply run:

git kommit

and follow the prompts to either:

  • generate a new commit message,
  • edit the generated commit message in your editor such as vim (can be selected by passing --edit flag),
  • proceed to create a commit with the generated commit message (can be selected by passing --approve flag),
  • or quit.

Relevant context can also be passed with --message (or -m for short). For example, a usage like this:

git kommit -m "performance increased by 10x with these changes"

will create a commit message that looks like:

perf: increase performance by 10x
 
- Add ...
- Remove ...
 
[Generated by Kommit]

What I’m particularly happy about is the cost subcommand:

git kommit cost

which would print something like:

💰 Kommit Financial Therapy Session 💰
 
Repository                          Cost ($)
 ~/github.com/username/project-1     0.00178
 ~/github.com/username/project-2     0.00064
 ~/github.com/username/project-3     0.00090
 ~/github.com/username/project-4     0.00047
 ~/github.com/username/project-5     0.00968
 ~/gitlab.com/username/project-6     0.00144
 ~/gitlab.com/username/project-7     0.00500
 ~/github.com/username/project-8     0.00058
 TOTAL                               0.02049
 
 
  Use ↑/↓ or k/j to navigate
  Press Ctrl+c or q to exit

Potential Improvements

Here are a few ideas that come to mind for future improvements:

  • Increase diff context: pass more lines of surrounding code to the LLM using something like git diff --staged -U10 to help the model understand changes more accurately.
  • Enrich prompts with metadata: include file types, commit history (relating to the changes), or project specific-cues.
    • Here lies a challenge I come across often, which is evaluation. Without persisting the generated output somewhere, it’s difficult to track regressions for quality.
  • Support other commit conventions: allow users to choose from different commit styles such as gitmoji.

What I Learned

I learned that any executable files prefixed with git- automatically becomes an extension of git so building this binary as git-kommit makes kommit available as a native Git command.

I also noticed a substantially increased baseline for commit message quality, especially from folks who usually treat them as an afterthought. 😅

Using kommit shifted my focus more towards logically grouping similar files or changes together, and letting the tool handle the rest.

I’ve heard great things about Claude Code, and it might make kommit obsolete. But building it was a fun and rewarding weekend project. One thing I’d like to try with claude is making a large set of changes and letting it generate multiple logically flowing commits, effectively removing the stating step from the kommit workflow.