The prompt that wasn't plugged in

The nodes app has a name. Julia’s route: the founding example was learning the lore of Marvel, lore traces back to folklore, and the wordplay assembles itself, Lore + na + More: LorenaMor, which a Spanish ear hears as “Lorena, amor”, with moor, the unexplored wilderness, hiding in the surname ▸ 3:17. A name born the house way: brainstorm until a word carries three meanings at once, then stop.

The day’s engineering lesson is better than the naming. Wiring the AI in, the responses came back as walls of text, so he tightened the prompt: answer in one paragraph, max four lines. Nothing. Tried Gemini 2.5 Flash, Flash-Lite, Pro, “no me está funcionando ni el 2.5 Flash ni el Pro, ¿qué está pasando ahí?” ▸ 6:19. Three models, apparently all disobedient. Then he pasted the code into GPT: the system instruction was never being sent ▸ 6:34. “Estaba enviándole los mensajes sin la instrucción.”

antes de culpar al modelo, verifica que el prompt viaje →

The debugging heuristic deserves its frame: when an LLM seems to ignore instructions, the boring hypothesis, the instructions aren’t arriving, outranks the exciting one, the model is misbehaving. Prompt engineering has a plumbing layer, and a prompt edited with great care in a file that never reaches the API is indistinguishable, from the outside, from a stubborn model. One console.log of the actual request would have saved three model swaps.

With the wire connected, the loop works on camera: ask about Hawking radiation, get a tight answer fast, tap a concept, watch a new node bud ▸ 4:51. Lorena is alive, and she finally does what she’s told…