The Dark Arts of Skill Engineering — Paul Bakaus, Renaissance Geek

Paul Bakaus once turned the entire web orange. He wrote jQuery UI, shipped an orange default theme, assumed people would change it, and watched them not. So when he points out that the purple gradients everyone learned to mock came from a CSS framework's default sample page, and that AI generated design now has its own house shade of beige, he speaks as the cause rather than the critic. That history is what makes his central claim land. A ban does not produce originality, it relocates the model one cluster over: forbid one overused font and it reaches for the next nearest thing in latent space. Slop is a moving target. The median is the model's gravity, and a few hundred lines of carefully written prose cannot pull against it. So he stopped treating a skill as a packaged prompt and started treating it as an extension of the harness, which is what this workshop is really about. Nine techniques, each aimed at something prose cannot do. Two sub agents kept blind to each other, one playing design director and one running a deterministic linter, because a single thread grading its own work always says it did well. Naming your top three fonts and then throwing all three away to shave off the safe prediction. A script whose only job is to hand the model a random seed it could not have guessed. Instructions emitted on standard output rather than buried in prose, which models follow far more reliably, at the cost of prompt caching. Hooks that block a write instead of complaining afterward. And the discipline underneath all of it: if a gate can be skipped, it will be. Speaker info: - https://x.com/pbakaus - https://linkedin.com/in/paulbakaus - https://www.paulbakaus.com/ Timestamps: 0:00 - Impeccable, and why it started 4:57 - Why banning a font does not work 6:25 - Prompting is the floor, harness engineering is the ceiling 7:55 - One: make it argue 10:11 - Two blind sub agents, one synthesis 16:59 - Two: forcing divergence 20:38 - Three: routing like a mixture of experts 22:47 - Four: giving a skill memory 25:22 - Five: scripts that talk back 29:36 - Six: hooks that fight back 34:37 - Seven: live wiring the browser 40:04 - Eight: it worked on my machine 44:04 - Every model is overfitted differently 47:20 - Nine: design for the weakest model 48:26 - If a gate can be skipped, it will be 51:59 - Inside the eval harness 56:18 - Can taste be evaluated? 58:27 - The future of skills 1:02:20 - Packaging and distribution