Blog
Notes on compounding leverage
Latest run
Staying current is a practice, not an event
Learning a language or learning to code was a project with an end. Keeping up with agentic tooling isn't — it's the same self-teaching method, run continuously, with no finish line in sight.
Read the post →Two threads
-
Harness engineering
Designing agentic coding harnesses, writing reusable skills instead of one giant prompt, and running coding agents unattended.
-
Self-directed learning
How self-teaching actually works, from picking up a language solo to staying current in a field that reinvents its tooling every few months.
Earlier runs
-
The highest-leverage productivity change was a data plan
I spent years trying to will myself into a consistent learning habit. The thing that actually worked wasn't willpower — it was upgrading a mobile data plan so commute time stopped being wasted time.
-
Pick your posture before you pick your prompt
Why an agentic harness needs an explicit operating mode — a fast autonomous posture and a slower high-ceremony one — instead of a single prompt trying to serve both.
-
Running agents unattended, overnight, and not regretting it
Recoverability beats approval gates when nobody is watching the agent work — a practical set of rules for safe unattended runs.
-
The method was never about the language
I taught myself two languages before I taught myself to code. The thing that transferred wasn't vocabulary or syntax — it was a method for turning 'I don't know this' into a plan.
-
Composable skills beat one giant prompt
Why small, reusable playbooks scale better than a single monolithic system prompt for an agentic coding harness.