Research Radar

The Problem

The product operating model is shifting

AI is rewriting product management and development. The roles are shifting, the handoffs are shifting, and the whole product operating model needs another look.

Most companies that take product work seriously have settled on the **product trio**: a PM, a designer, and a developer working together to make sure viability, desirability, and feasibility all get a vote. The shorthand for this is **POM** — Product Operating Model. It's a good model. It still works.

But it stops being the only model the moment modern AI tools come into the picture. Agents start doing chunks of the trio's work. The handoffs change shape. The question of what a human does versus what an agent does becomes a real design choice, not a hypothetical.

I'm curious how different teams and companies are actually answering that right now. Who's keeping the trio intact? Who's shrinking it? Where are the agents taking work, and where are they being kept out? What does a handoff look like when one of the participants doesn't have a calendar?

To follow that shift without spending half my week on it, I needed someone scanning the field for me — reading the articles, watching the YouTube talks, tracking what frontier teams publish. That's what Research Radar does: it fetches, filters, and curates a list I'll actually read.

The Solution

A scout that runs Tuesday and Friday at 7:00

Twice a week, while I'm having coffee, a Claude agent runs on my Mac. It scans the past 7 days of content from a named list of Substacks and runs broader web searches around AI-assisted product management, discovery, user research, prototyping, and validation. It skips hype, generic AI takes, ML deep-dives, and vibe-coding tutorials that aren't tied to product workflow.

What it produces:

**Notion database rows** — one per finding, with title, source, summary, link, and tags
**A digest email to my inbox** — subject line, top items, brief notes per finding, plain text so it reads fast on a phone

Notion is the source of truth. The email is the trigger that gets me to look. If something matters enough that I want to act on it, I open Notion and link the article to whatever piece of work it informs.

macOS launchd handles the schedule with one plist holding two cron entries. A bash wrapper sets up the environment, clears the previous brief file, and invokes Claude in headless mode with a narrow allowlist of MCP tools — only what the skill needs.

The Claude skill (`SKILL.md`) is the actual intelligence. It runs parallel web searches across the named sources, fetches and reads candidate articles, filters by scope, writes findings to Notion, and writes a brief file. The bash wrapper then sends that brief via a small Python smtplib script.

Logs go to `~/Library/Logs/research-radar.log` so I can see what happened on any given run.

Why this works

Curation, not aggregation

An RSS reader gives me 50 titles. The radar gives me 5 to 8 findings that are actually relevant, with summaries based on what's in the article (Claude can fetch and read), not just the headline.

The scope is enforced in the skill prompt itself:

7-day window only — anything older is skipped silently
Topics in scope are listed explicitly; out-of-scope topics are listed too
Named Substacks searched site by site, not via "best of" aggregators

The Notion database has been quietly accumulating since the radar started running. It's now a searchable archive of what's been published recently in AI product work, tagged and linkable to whatever I'm writing or pitching.