@mattpocockuk: Teach Techniques That Work in Any Harness

@mattpocockuk offers an alternate data point to talk of courses dying: this has been his best year selling self-paced video courses. They have been 98% of his income since 2022.

What is worth learning with AI, he argues, are techniques that are harness-agnostic, model-agnostic and language-agnostic, and that raise the agent’s output quality. He has explored software fundamentals, module design and test quality. He sees room in user interfaces, company process, leadership, observability and product design.

He would not make his Total TypeScript course now, he writes, because sales have “dropped off a cliff.” It leans on APIs and language features that are easy to ask agents about. He would make “TypeScript for AI” instead.

In a separate post he sketches a tool that checks deterministically whether a PR can be automerged. His proposed constraints: the PR only reorganizes code, keeps the module graph as it is and keeps test coverage from getting worse. He says it would not work for new features.