Writing
Things the work taught me.
The builds keep teaching me things I did not go looking for. These are the ideas that survived contact with the build.
-
Where the model stops
Find the small part of the work that needs a model, and keep everything else with code, people, and evidence.
-
Measuring Context Costs in an Orchestrator-and-Subagent Workflow
Three months of published context rules measured on one repository. The fan-out got cheap and the cost moved to the main thread.
-
Measure the Assembly Before You Blame the Model
Context budgets, retrieval chains, and choosing an inspectable level of autonomy.
-
The PM work a model can do
A one-day text-to-speech build where measuring a perceived audio flaw changed the fix.
-
Smarter Every Time It Fails
6,245 eval runs since June 10, 844 below the pass bar, each one read, diagnosed, and turned into a permanent regression case.
-
What the Model Sees
Context engineering and system design, checked against a local-first agent harness: token budgets, retrieval failure modes, and the autonomy ladder.
-
The 300 Second Write
What a wedged GPU taught me about picking models, measuring launches, and keeping an AI product alive once real work depends on it.
-
Three Scouts in 47 Milliseconds
When to fan a task out across cheap parallel agents, when to keep it on one thread, and what a 181,083 character failure loop taught me about the seams in between.
-
An adversarial review gate that caught 4 leaks across 5 drafts.
An adversarial agent has to clear every draft before a person signs off. What it caught, and what that changed about how drafts get written.
-
The screen is not where finished things come from.
What changed when my children stopped asking what they could play and started asking what they could make.
-
Build enough things to see the patterns.
I can make almost any idea sound coherent in a document. Software is less polite. Building shows me which parts of an idea are actually true.
-
A model forgets what it drew one image ago.
Ask for the same rider twice and you get back a cousin. The fix became a rule about where variation gets to live in a system.
-
I took away my smartest AI on purpose.
If the smartest model in the chair disappeared tomorrow, how much of what I built would still stand? The answer changes how you build everything.
-
Research first, then compress it into the system.
Before the arcade got built I broke down how the winners do it. Then I compressed what I learned into playbooks the next build actually reads.
-
We built an arcade this weekend.
Building a home arcade of tiny games with my kids, and what the kids taught the build.