Make it ask you questions

The trigger is self-defense: TikTok, “pura dopamina… al final no aprendes nada” ▸ 0:31, so he asks Cline to block it, and appends the clause he now uses on every task: “hágame la mayor cantidad de preguntas posibles para entender mejor lo que quiero hacer” ▸ 10:07.

The philosophy behind the clause is the humbling one: “cuando la IA no hace una tarea correcta, el problema no es de la IA, el problema es de uno” ▸ 10:46. His taxonomy of failures reduces to one cause wearing two masks: the model lacked information, either because its training predates your library, or because you explained yourself badly, which “termina siendo información incompleta” too ▸ 11:43. The questions clause attacks both at once by making the model surface its ambiguities before spending effort.

las preguntas del modelo son tu spec, redactada por la parte que va a ejecutar →

The block itself proves the point live. First question back: block on this computer only, or across the whole network router? “Yo fui como que: wow”, because the vague version might have firewalled the household ▸ 12:10. Second: what’s the goal, productivity, parental controls, privacy?, each answer steering toward a different implementation entirely ▸ 13:00. With the spec extracted, the fix is three lines in C:\Windows\System32\drivers\etc\hosts and admin rights ▸ 14:25, and TikTok now returns a clean refusal. It’s the Mermaid-diagram principle in conversational form: the cheapest place to catch a wrong assumption is before the work, and the model is perfectly capable of interviewing you, if you order it to…