← all posts site ↗

use the ai as an accelerant, not a coauthor

figured out the model for how i actually want to use ai after a long conversation about why building feels so much better than being handed the build. the working model is: write a very detailed plan myself, down to the nuances of every system, then let the agent one shot as close to the whole thing as it can, and then i step in and own the finish and the small tweaks. the first pass is surprisingly good at low complexity and degrades as complexity grows, so the division is clean. the agent gets the architecture and the mechanics, i get the judgment and the finish pass.

the insight that made it click is the throwaway problem. when the ai writes the code from scratch i don't care about it, because i didn't author it, so it's disposable to me no matter how good it is. the fix is to move where the authorship happens. when i write the plan, i become the author of the design, and the code becomes a translation of my intent, so the vested interest lives in the spec and in the finish pass i do myself. but the plan can't lock everything, if the spec decides every detail then the interesting choices were all made by the plan, and i still don't own it. thorough on architecture and the hard parts, deliberately open where i'll take over.

and i pushed back on the idea that taste is the last scarce skill. taste is reproducible, an ai can copy anyone's surface taste. what actually survives is end to end system comprehension, knowing why every manager exists, what the json drives, why the camera does what it does, and that's exactly what the sixty to hundred gap demands. the value of the work is moving, not disappearing. ai devalues outputs, it makes them cheap, but it can't devalue the experience of doing the work because that has no price and i never sold the output anyway.