I Built An AI Git Commit Message Generator
I used to write bad commit messages like "fixed bug" or "updated stuff." This made my pull requests messy. I decided to automate it using AI. I wanted to generate clean, conventional commit messages from my code diffs.
It was harder than I thought. A simple prompt did not work.
The early attempts failed for many reasons:
- The AI wrote long paragraphs instead of short lines.
- It ignored the required "feat:" or "fix:" prefixes.
- It hallucinated features that did not exist in the code.
- Local models were too slow and wrote in a poetic, useless style.
I eventually built a hybrid system that works. Here is the structure I use:
- Type Classification: I ask the model to pick a type like feat, fix, or chore before writing the message.
- Context Truncation: I only send the first 250 lines of the diff. This saves money and keeps the focus sharp.
- Validation: I use a regex to check the output. If the message is wrong, the script tries again.
- Low Temperature: I set the temperature to 0.2. This makes the output consistent and boring. Boring is good for logs.
I do not use automatic commit hooks. AI makes mistakes. I run the script as a git alias. It suggests a message, and I review it before I commit. Manual review is the best insurance against a bad git history.
Three lessons I learned:
- Token limits are your enemy. Large diffs will break your budget or your prompt.
- Validation is mandatory. Without it, you will get nonsense.
- Speed matters. I use GPT-4o-mini because it is fast and cheap.
If you work on sensitive code, do not send your diffs to an external API. Use a local model instead.
How do you handle your commit messages? Do you use AI or write them by hand?