Three whiteboards of my own setup
My AI group chat tried Claude’s new whiteboards on our own setups. The tool routing board was the useful one, for checking my routing against the mental model I had.
Nov 11, 2026Moving an agent harness to a new job
I kept my agent setup when I changed jobs and stripped the old company out of it. The inventory caught three things that would have broken quietly.
Oct 11, 2026What would follow me to a new job
A thought experiment before I changed jobs: what would follow me? Not the folders. The memory, and ten scheduled jobs still writing company material.
Sep 20, 2026My agent was never on the model I thought it was
Step 16 of my own migration guide said to trust the session metadata. I did, for three weeks. The app’s log said Opus 5 on every one of 168 turns.
Sep 1, 2026Move Dispatch to Opus 5 without losing your setup
There is no model selector on Dispatch, so deleting the conversation is the only way to upgrade. Whether you lose anything was decided months earlier.
Jul 26, 2026A VP said it shipped. That’s not a source.
An agent turned one sentence from a meeting into a sourced claim. The claim survived verification — and the receipts came with intel no summary had mentioned.
Jul 23, 2026My always-on agent almost never pinged me. The budget wasn’t the problem.
Three scheduled scans a day and almost no pings. The diagnosis: the message budget had never once bound, and a denser cadence made the system quieter.
Jul 23, 2026Six agents critiqued my card design and killed my favorite idea
I ran a six-voice critique council of AI agents to take a social card from sixteen rough directions to a shipped system. The full session, on video.
Jul 23, 2026Evaluating new tools against your own graveyard
Eight agents reviewed five GitHub repos — five reading source, two reading a year of my own logs to decide what would actually survive here.
Jul 19, 2026A memory system had been quietly wrong about me since March
The corrections landed in the graph. The profile derived from it never rebuilt. Then the cleanup that killed it reported clean from scans that never ran.
Jul 15, 2026Two agents shared a filesystem for weeks without knowing the other existed
Two agents, one vault, weeks of parallel work. The bridge design was right — but the fixes were landing in mirror files. One docs question caught it.
Jul 5, 2026Where knowledge goes to die: making an agent audit its own memory
I pointed Fable 5 at its own memory system and asked what actually compounds. It found the graveyards, indicted its own machinery, and rebuilt it in a session.
Jul 4, 2026The Year-in-Review That Corrected Itself
I shipped a clean analysis of a year of my own posts — then pulled the real numbers and watched the headline reverse.
Jun 30, 2026Redesigning a Branded Deck Through an Agent — and What Its Renders Missed
An open-source tool let an AI agent edit a branded deck in place. The renders looked clean — the round-trip back through the real slide software told a different story.
Jun 26, 2026Reorganizing a Knowledge Base by Lineage, Not by Filename
A knowledge base that had drifted into two competing folder schemes, reorganized by tracing each file to the work that made it — and the distinction a bulk find-and-replace couldn’t see.
Jun 17, 2026Wiring a Second Opinion Into How I Plan With AI
I used to paste every plan into a second model for review by hand. Automating it meant teaching the system what I keep rejecting.
Jun 17, 2026Nobody was watching that page
A page nobody was watching claimed the product could do something it cannot. The only thing that surfaced it was a refund.
Jun 4, 2026Level 10, Level 3
A colleague stopped my demo four sentences in. Knowing my own setup cold is what made me the worst person to explain it.
Jun 2, 2026The Skill That Only Ran on My Machine
Publishing one of my own AI skills meant proving it could run with none of the infrastructure it was built on.
Jun 2, 2026New entries are added as they’re elevated from my drafts queue. Follow along on X where I share highlights.