Field notes for people building useful AI systems.
Essays on agents, prompts, product judgment, and the operating habits that make AI work less fragile.
Browse the concept indexUse ChatGPT agents like a careful operator
ChatGPT agents are useful when you give them bounded work, the minimum access, and a visible checkpoint before anything changes. Treat them like a careful junior operator, not magic.
Evals as contracts: how to know your AI system is working
Stop thinking of evals as tests. Start thinking of them as contracts: a one-page agreement that names the failures, the thresholds, and what happens automatically when the system breaks it…
Why your AI pilot is stuck (and what to do about it)
Most stuck pilots are not stuck on the model. They are stuck on one of three things: nobody owns the pipeline, nobody has defined good enough, or nobody has told the money a story it can a…
The three prompt patterns I actually use in production
After shipping AI in four companies where wrong answers cost real money, I stopped reaching for clever prompt engineering and started reaching for three patterns. Here they are, with copy-…
How to build your first AI agent in an afternoon
A working plan for consultants and product people who want to go from zero to a running agent by five o'clock: one task, hardcoded data, a loop you control, and a stop point. No buzzwords.
Working on something like this? Start a conversation