What we shipped, what broke, and what we changed our minds about. Written for operators, not the timeline.
I built a research skill the modern way: handed my expertise to an LLM and layered review tools on top. It looked rigorous, until I verified the gold-standard example I was shipping and found it broken. Here's what rebuilding the gate taught me about whether you can trust your own evaluation.
Read article →I spent a month building a full-stack application: server, API, 11 pipeline phases, 21,000 lines of code. The thing that shipped was 43 files in a folder smaller than a hero image.
Read article →We built one discovery pipeline on the Agent SDK. Then we built the same thing in Claude Code. Here’s what both approaches actually require — tested across two client engagements.
Read article →When the process is known, fixed workflows beat autonomy. Here are the guardrails we now use.
Read article →A repeatable 3-flow system for collecting updates, tracking status, and chasing non-responders automatically.
Read article →We run 4–8 week Workflow Builds with operators who want their AI to actually run in production. Bring the messiest workflow you have.