Writing
Notes from inside the method.
What breaks in vibe-coded codebases, why a passing suite is not evidence, and what it actually takes to grade work a model produced. Written from runs on real repositories, with the numbers attached.
Nothing here is published yet. The shelf is listed so you can see what is coming and tell me which one you want first, rather than arriving at an empty page.
A green test suite is a claim about the tests
Six guards in one of my own repositories were passing and structurally incapable of failing. Here is how each one got that way, and the check that finds them.
The dangerous column is the one nobody measured
Broken work is visible and gets fixed. Work nobody ever checked either way looks identical to work that passed, and it is where a rebuild runs aground.
Why the model that wrote it cannot be the model that grades it
Self-verification is the cheapest thing in an agent pipeline and the least load-bearing. What changes when the grader comes from a different family and never saw the plan.
What vibe coding actually leaves behind
Not bad code. Copied logic, boundaries nobody enforced, and a happy path that became the specification. The five patterns that show up in nearly every prototype.
Reading a coverage number you did not produce
Two of my repositories carried a headline figure that no run had ever printed. Numbers get copied forward; the run that made them does not.
Forty-five products, one person, and the honest accounting
What an agent fleet actually costs, where it saves nothing, and the three places a human still has to sit in the loop or the whole thing quietly degrades.
Want one of these sooner?
Say which and it moves up the list. If you have a repository that would make a better example than mine, that is a different conversation and a more useful one.
Get in touch