Prompting Commits Instead of PRDs

Kanban Board

When I take on a complex task with AI, I used to reach for detailed documents: a PRD-style write-up, maybe a .cursor doc, or some long prompt stuffed with context. It worked, but it always felt heavy. Too much ceremony up front, too much assumption about what I already knew.

Lately I’ve been trying something lighter: prompting commits.

The Approach

Here’s how it works:

  1. Set the goal
    Start with the big picture outcome. Something like:
    “Build an image editor with upload, crop, and filters.”

  2. Ask for a commit plan
    Instead of writing a spec, I have the AI propose a series of commit-lint style commits that would logically arrive at that goal. Example:

    • feat: add image upload component
    • feat: integrate cropper tool
    • feat: add filter options with preview
    • chore: refactor state management
    • fix: prevent double uploads

  3. Orchestrate the work
    I feed that list to an orchestrator, which gives each commit to an agent. Each one becomes a focused, self-contained task with a clear output. The result is a set of clean, atomic commits that map directly to progress.

Why It Works

  • Laser focus
    Each commit is a small, defined step. No sprawling prompts, no drifting scope.

  • Natural granularity
    Commits are already the unit of work we use in development. Using them as prompts feels intuitive and portable.

  • Better outputs
    Agents don’t lose track of the bigger goal, but they also aren’t drowning in it. They’re solving one step at a time.

  • Document alternative
    This sidesteps writing big PRDs or meta-docs. Instead of describing every detail up front, you bootstrap the process with a commit plan and let the work unfold in code.

Not a Silver Bullet

This isn’t a solved problem or a one-size-fits-all process. Sometimes a traditional spec makes sense. Sometimes you need broader architectural notes. Prompting commits is just another tool—one that’s worked surprisingly well for me when I want momentum without overhead.

The Bigger Picture

What excites me here is less about commits and more about the pattern: using structures we already trust (like commit messages) as scaffolding for AI workflows. It’s simple, familiar, and immediately actionable.

We’re all still experimenting with how to get the most out of AI-assisted development. For me, prompting commits has been one of those small, helpful hacks—an alternative starting point that makes complex tasks feel concrete and approachable.