Principles

Practices I've stolen, tested, and kept. Take whatever works. Ignore the rest. If one of these is wrong for your situation, it's wrong for your situation — don't argue with the page, argue with your own next commit.

The failing test is the lever

When the codebase is in trouble, don't reach for the cleverest abstraction. Reach for a failing test. Watch it go red. Write the smallest thing that turns it green. Refactor. Repeat.

A bug that isn't pinned by a test is a rumor. A feature without a test is an intention. The test is what converts either one into something the codebase actually has to honor — every commit, forever.

Takeaway: red → green → refactor. If you're stuck, write the test first. The test is the lever you pull when nothing else moves.

Own your weights

Run models on your own metal whenever the task is small enough to fit. Local LLMs are not always smarter — they are yours. There is a difference between renting intelligence by the token and housing it on a disk next to your dev machine. Have both, but never give up the second.

Frontier models for the hard problems. Local models for the loop — the boring, repeated, all-day work that adds up to a practice you don't have to apologize for.

Takeaway: if a task you do daily can be done by a 7B model on your laptop, move it there. The token bill is not a feature.

Manifestation as scope-setting

Manifestation isn't woo. It's the discipline of naming the thing precisely enough that you'd recognize it if it walked in the room. If you can't write the acceptance criteria, you can't manifest the feature. If you can't write the acceptance criteria for the life, you can't manifest the life. Same engine.

write the outcome down in a single sentence. Read it back tomorrow. If you don't recognize it, sharpen it.