I’ve always been meticulous about commit messages—seeing them as a crucial part of the project’s story—but writing good ones felt like a constant context switch. Six months ago, I decided to offload that entire task to an AI coding assistant for every single commit. The goal wasn't just to save time, but to see if consistent, AI-generated messages could actually improve our team’s repo history.
Here’s my setup and what I learned:
**My Workflow Recipe:**
* I use a CLI tool that hooks into `git commit`. It takes the diff and feeds it to the AI with a very specific prompt.
* The key was crafting an instruction set that matches our team’s conventions. I don’t just ask for "a commit message." I specify:
* Use the imperative mood ("Fix crash" not "Fixed crash").
* A subject line under 50 chars.
* A body that separates *what* changed from the *why*, especially for non-obvious fixes.
* Reference relevant issue tracker IDs.
* **Crucially:** Highlight any changes related to crash resolution, user-facing behavior, or beta feature flags—my personal areas of focus.
**The Surprising Results:**
* **Consistency is the biggest win.** Our main branch history is now uniformly formatted, which makes scanning through `git log` or using `git blame` much clearer.
* **It catches details I’d gloss over.** When I’m deep in a refactor, I might forget to note a specific edge case I handled. The AI, working from the full diff, almost always includes it.
* **The downside? Sometimes it’s *too* generic.** Early on, I’d get messages like "Update code" for complex changes. I had to refine my prompt to demand specifics from the diff itself (e.g., "mention the modified function names").
* **You must review.** I never just accept and push. I treat the AI’s suggestion as a first draft, ensuring it accurately captures the intent, especially for UX tweaks or beta tooling changes. This review takes seconds, not minutes.
After half a year, I wouldn’t go back. It’s like having a dedicated documentarian for my code changes, freeing me up to focus on the actual logic and edge cases (which, let’s be honest, is the fun part). The commit history has genuinely become more useful for my teammates and future me.
Has anyone else tried automating this part of their workflow? I’d love to compare prompt strategies!
happy testing!
edge cases matter