Run PRD-driven coding agents

The agent workflow gap

AI coding agents can implement a lot of work quickly, but they still need structure. Without a loop, a developer has to manually decide the next task, rewrite context, check progress, run review, and clean up state.

A PRD-driven workflow solves part of the problem by defining what should be built. Hal solves the loop around it.

How Hal structures the work

Hal stores project workflow state in .hal/, reads product requirements, and runs implementation iterations against structured user stories. It can generate or convert PRDs, validate them, run the coding loop, and update progress as stories complete.

Each iteration has a narrow target. That keeps the agent focused and makes the result easier to review.

A practical workflow

1. Initialize Hal in the repository.
2. Capture the feature as a PRD.
3. Convert and validate the PRD.
4. Run a bounded number of implementation iterations.
5. Review, test, report, and archive the completed feature state.

For teams that want more automation, hal auto can run a fuller pipeline from analysis through archive.

Why fresh context matters

Fresh context reduces the chance that an agent carries old assumptions into the next story. Hal keeps persistent state in files while giving each implementation pass a cleaner working window.

When to use this workflow

Use PRD-driven coding agents when the feature has multiple steps, when you need an audit trail, when you want CI and review gates, or when a solo founder needs to run development loops without acting as a full-time project manager.