NeuroDiscovery AI · 2026 · Internal · Agentic PM tooling

PM delivery system on Claude Code

Claude Code skills that turn meeting notes and chat into journey tickets, prototypes inside the real codebase, handoffs, tests, demos and release notes, with a human gate at each step.

My roleDesigned and built it; I run my product work through it every day.

2 wks → <1 wkidea to releaseMy figure

304 / 304product tests passing on prototypes built in the real appRun log

0 → 4 of 4ticket eval after gating; 5 leaked identifiers down to 0Synthetic test note

Anurag from behind facing a wall screen with a glowing orb — The desk the system runs from. AI generated still.

The challenge

A small product team running six products: thin tickets, context lost in chat, prototypes drifting from the real interface.

What shipped

14 skills and a local dashboard, with ticket evals, routing evals and full test suites as gates.

What moved

Idea to release went from a two week sprint to under a week, with 304 of 304 product tests passing on every prototype

Notes become tickets written as user journeys. Tickets become prototypes inside the real product repository, on a branch, using the real components. Nothing reaches Jira or an engineer without my approval, and every correction I make is written back into the skill so it does not repeat.

How it was evaluated

Ticket eval: testable acceptance criteria and zero personal data
Routing eval on a golden set of intakes
Build, types and the full test suite on every prototype

5 / 5intakes routed correctly after one fixVerified record · Small golden set

What I learned

Prototypes belong in the real codebase.
A correction that appears twice is a defect in the skill.