Hal: autonomous AI coding loops for PRD-driven development

The problem

AI coding agents are powerful, but long sessions can drift. Context gets noisy, product intent gets lost, and progress is hard to audit. A solo developer needs the speed of agents without giving up structure.

PRDs help define the work, but a PRD alone does not run the loop. Someone still needs to break work into iterations, keep state, validate output, review changes, run CI, and decide what happens next.

What Hal does

Hal is a CLI for autonomous PRD-driven development. Feed it product requirements and it runs coding agents one story at a time. Each iteration gets a fresh context window, applies project standards, implements the next story, commits changes, and updates progress.

Hal supports both a manual workflow and a fuller auto pipeline.

Manual workflow

A typical manual flow is:

1. Run hal init to create the project workflow files.
2. Run hal plan to generate or capture a PRD.
3. Run hal convert to turn markdown requirements into structured runtime state.
4. Run hal validate to check quality.
5. Run hal run to execute implementation iterations.

Auto pipeline

Hal also includes a deterministic auto pipeline for larger loops. The pipeline can analyze reports, create a spec, branch, convert, validate, run, review, check CI, report, and archive state.

That makes it useful for repeated agent-driven development cycles where every run should leave behind inspectable state and artifacts.

Engine support

Hal is designed around pluggable AI coding engines. It can work with Codex, Claude Code, or Pi, letting a project keep the same PRD-driven workflow while changing the underlying agent.

Best use cases

Hal is useful for solo developers and technical founders who want agents to implement product work from clear requirements, keep iterations small, preserve project standards, and avoid one giant context window.