Hey friends, I know a Friday afternoon email is bad for ✨engagement✨ but I just remembered I haven't sent one of these out an about a month, so it's going to have to do. The roundup is still heavy on the AI stuff, I know, but I'm still very much enjoying how the PM role is changing, so I expect to be locked in (that's what we say now I think?) for a while longer.

Anyway, have good weekend, ok? -Rian

Full Posts

From Assistant to Collaborator: How My AI Second Brain Grew Up

The fifth post in the AI second brain series, covering the shift from "tool I invoke" to "system I dispatch." Three capabilities drove the change: multi-agent workflows that run specialist agents in parallel and synthesize their findings, a four-layer cross-session memory system (stable facts, session log, corrections, decisions), and domain skills that encode institutional knowledge about specific products and their failure modes. The post includes a concrete walkthrough of the customer escalation investigation workflow — five phases from ticket classification through blind validation — and four practical observations about the cost of maintaining it: process design over prompt writing, validation built into the workflow, the discipline of actually running /session-end, and the ongoing drift of domain skills as products change.

I am finally — FINALLY — off WordPress

After 18 years on WordPress and Dreamhost, elezea.com is now running on Astro, deployed to Cloudflare Workers, with no database and no server-side runtime. The migration that kept getting filed under "someday" only became possible after iterating on a detailed plan with Claude Code over several weeks — the plan eventually ran to 1,300 lines and covered everything from shortcode handling to DNS cutover. The actual migration took a single weekend. Search runs client-side via Pagefind, content lives as Markdown in a git repo, and the site is the fastest it's ever been. The broader point: AI is good at making vague, scary projects legible — iterating on a plan long enough turns an impossible thing into a checklist.

Link Posts

• Agentic Manual Testing — Simon Willison on two practical testing habits worth adopting: write demo files to /tmp to avoid accidental commits, and use red/green TDD when an agent finds a bug so the fix lands in permanent automated tests.

• Negative Space in Writing — Tracy Durnell on what gets lost when writing formats optimize for skimmability: the gaps where readers build their own meaning, follow a hunch, or arrive at a conclusion themselves.

• Endgame for the Open Web — Anil Dash's essay is worth reading for one definition alone: the open web is the radical ability to create and share using open specs, free platforms, and no gatekeepers. Every part of that architecture is now under pressure.

• Eight Years of Wanting, Three Months of Building with AI — Lalit Maganti on building a SQLite parser with AI after eight years of procrastination, including an uncomfortably accurate comparison of "just one more prompt" to slot machine addiction.

• AI Might Actually Need More PMs — Anthropic's Amol Avasare argues that AI is raising engineering leverage faster than PM leverage, which makes PMs who can sharpen priorities and decisions more valuable, not less.

• Zombie Flow — Derek Thompson traces the "flow" concept from Csikszentmihalyi to algorithmic feeds that simulate absorption without challenge, achievement, or volition, and argues the real skill now is learning to escape the fake version.

• No One Else Can Speak the Words on Your Lips — Ben Roy on why real writing is bottom-up discovery: you don't know what you want to say until you've written it, and LLMs are structurally incapable of that process.

• Evals Are the New PRD — Braintrust's argument that for AI products, the eval replaces the PRD: it defines what "good" looks like, sets the acceptance criteria, and gives engineering a clear directive: "make this number go up."