Spec-driven workflow for AI coding agents
Addressed to
Anyone who has watched an AI agent confidently build the wrong thing. You approve a plan, it becomes phase documents in your repo, and the agent works to those documents one phase at a time.
/plan2code-0-pathfinder
For an idea that arrives too foggy to plan. It charts the idea as a map of decision questions and clears exactly one per session until nothing is left to decide.
it never answers its own questions — that's the point
/plan2code-1-plan
A real back-and-forth about what gets built. Requirements pulled apart, your codebase actually read, the stack confirmed rather than assumed.
/plan2code-2-document
The plan stops being a conversation and becomes files — an overview plus one self-contained phase document per phase, cut into one-story-point tasks.
/plan2code-3-implement
Build to the drawing, one phase per conversation. Planning context bleeding into implementation is where most agent drift starts.
reply “approved” — nothing advances on a guess
/plan2code-review
An independent senior engineer with fresh eyes and no attachment to the code that just shipped. It reads every changed file completely and reports only what it can prove — file, line, and a concrete fix.
/plan2code-4-finalize
Close it out. Every task audited against the spec it came from, a summary of what was built, docs checked for drift, and the folder filed away.
A fresh conversation for each step, and each implementation phase.
No code until the plan hits 90% confidence. The gate does not negotiate.
Checkboxes are the state, not the chat. Any session resumes cold.
Reply approved to close a phase. Nothing advances on a guess.
$ npx --allow-git=all git+https://github.com/jparkerweb/plan2code.git