How I work
The systems behind the output. The case studies show what got built, this page is the operating manual they share.
The shape of it
I don't use AI as a chat window. I build loops: systems that pick up work on a schedule, do it to a written standard, check their own output, and leave me a decision instead of a draft. The projects on this site are all the same five pieces arranged differently: automations, isolated workspaces, skills, sub-agents, and memory.
The rule that makes it work: agents research and draft; the opinion is mine. Nothing ships that I haven't put my own angle on.
Written standards, not repeated prompts
Everything I know how to do well gets written down as a skill, a playbook the agent reads every time, so quality doesn't drift. My reel engine carries my voice rules and a scoring rubric. Potential's agents carry a brand-intake doc that defines positioning once. Four of my skills are packaged and distributable.
When output gets rejected, the fix goes into the system, not the draft. The reel engine's hard gates, one-breath hooks, no manufactured first-person claims, each exist because something specific got through once and never should again.
Makers don't grade their own work
The single highest-leverage pattern I've found: separate the author from the judge. A drafter writes, a fresh-context critic scores against the rubric, a judge picks. Same-context self-review misses what fresh-context review catches, every single time.
One spine, many limbs
Notion is the shared state: content calendar, ideas inbox, activity logs the agents write to and read from. Postgres holds the numbers. Every night, agents snapshot live state to dated files nobody overwrites, that's what makes real week-over-week trends possible, and it's why a missed night is unrecoverable and worth engineering against.
The loop closes in public: Instagram for reach, this site for depth. Every guide that performs gets rewritten here in full.