AI Product Owner: Agentic Backlog Generation

What I built

Translating raw tickets into formatted user stories is repetitive, low-leverage PM work that the PRD estimates at 20–30% of the job. The target was a CLI that takes a ticket CSV and emits a backlog file of stories that need no human edit.

What I did

Built the tool on a ReAct loop with a reflexion pass so the agent critiques and rewrites its own output before returning it, and constrained the result with a Zod schema so every story is structurally INVEST-compliant rather than merely prose that resembles one. Added a --dry-run flag that prices a run before it happens, RICE scoring inside the prompt, and an Express wrapper so the same logic can be called as a service instead of a terminal command.

What came out of it