The pancake analogy
When AI fails you it's one of exactly two things, stale knowledge or an unclear ask, and the proof that the second is universal comes from dinner: 'deja la salsa para la otra' meant the pancake underneath to her and tomorrow's pancake to him, same words, two referents.
// trace: where this idea came from
- ↳ video diary @ 18:47 (las dos razones)
- ↳ video diary @ 20:51 (la panqueca de la discordia)
- ↳ Entry 113-1: Make it ask you questions (el protocolo de preguntas, ya descubierto antes)
The theory is stated with the confidence of someone who’s paid for it: models are now good enough that when they fail, it’s one of two reasons ▸ 18:47. Either the knowledge is stale, you’re on the newest Flutter and the model trained on one three versions back, fixable by handing it the docs ▸ 19:01; or you didn’t explain yourself clearly ▸ 19:27.
The second gets the proof from dinner. Four pancakes, stacked two per plate, guava sauce. As he reaches to finish the sauce on the top pancake, Julia says “deja para la otra” ▸ 20:51. He hears save the sauce for tomorrow’s pancake and objects; she meant the one underneath, on the same plate, right now ▸ 21:21. An argument ensues between two people who wanted the same thing ▸ 21:29. If a couple with years of shared context can collide on “la otra,” a model meeting your codebase cold never had a chance ▸ 21:48.
“la otra” tenía dos referentes; tu prompt también los tiene →
The protocol, refined from Severo’s questionnaire lesson into a habit: start every task in plan mode, have it read the files first, then “sácame todas las preguntas posibles que usted tenga” before touching anything ▸ 22:33. The step feels redundant exactly when it’s most needed: you’re sure the instruction is perfect, the questions come back, and “bueno, tienes razón, hay detallitos que yo no dije” ▸ 23:21. Ambiguity is invisible from inside the head that produced it; the questions are how you rent a second head…