AI you build by hand

Built as an Astro static site, content via Markdown content collections.

Fed the PRD to the model, generating structure and components in conversation.

A content home that accumulates over time, and a natural piece of evidence.

A reusable multi-agent pipeline — an orchestrator + five specialist agents (prd-writer / designer / coder / reviewer / tester); both this site and Show Radar were taken 0→1 with it.

spec.md as the single source of truth, injected on every cold-start handoff to stop context loss; judgment / taste set as three fixed human gates to catch errors upstream; cross-model review (Claude Code × Codex catching each other's bugs).

Turns AI coworking from "drop a goal and wait for a result" into a controllable "one orchestrator + a few human-in-the-loop checkpoints" flow; what it distills is the AI-era PM's core method — define constraints, design the collaboration protocol, make the call at the key gates.

Multi-source listings → scrape → LLM structuring → dedupe → preference scoring → summary → iOS.

Decouple scraping from extraction with fixtures, wire up the easiest source first, abstract the system as a "data pipeline" not "real-time monitoring", dare to cut the detail page, hybrid cloud/local architecture, write the PRD as a collaboration protocol for AI, and cross-review with multiple models (implement / find faults / re-check).

Proof that an AI-era PM can push a real need through AI coworking into a runnable, deployable, narratable data product. The most valuable output isn't the code — it's the hands-on method for how a PM manages information flow and quality control between agents.