Feed it the whole company
The job-application workshop becomes a worked example of context-window practice: dump the case PDF plus every page of the employer's site into Gemini, ask for the table that maps their branded services to the Google products underneath, and ground every claimed number in their own case studies.
// trace: where this idea came from
- ↳ video diary @ 19:06 (metió el caso y toda la página)
- ↳ video diary @ 20:11 (la tabla que desarma la marca)
- ↳ video diary @ 30:24 (los números salen de sus propios casos)
- ↳ Entry 119-1: Good enough is a stopping rule (la regla que casi rompe aquí)
The task from the Colombian company: a fictional client, XFood, retail logistics across four countries, data scattered in disconnected systems, pitch a solution on video ▸ 16:02. His method is the context window used as designed: into Gemini, in Google AI Studio because it’s free ▸ 18:08 (“GPT lo utilizo para consultas simples… ¿cuál es la capital de Francia?” ▸ 18:11), goes the five-page case PDF plus every tab of the employer’s website, products, services, and some fifty client case studies ▸ 19:06.
The sharpest move deconstructs the vendor itself: their services are Google products wearing brand names, so he asks for a comparison table, Xertica service on the left, the Google service probably running underneath on the right ▸ 20:11. Suddenly the jargon parses. The second-sharpest fixes his own draft: version one claimed improvements pulled from the air, “¿de dónde van a salir esos datos?”, so version two sources every number from the company’s own case studies, results they already achieved for similar clients ▸ 30:24. Same trick as citing a transcript: the claim inherits the credibility of its source.
desarma la marca en sus componentes; cita sus propios casos de vuelta →
The rest is honest tooling friction: Nano Banana draws handsome slides whose text says projeto ▸ 23:34 and refuses to fix them, the first batch comes out in English for a Peruvian client ▸ 22:36, and salvation is a Canva feature he’d never noticed that OCRs the image’s text, font and all, into editable layers ▸ 25:35. He considered shipping the typos and didn’t, “esto no se ve nada profesional” ▸ 26:58. Last touch, for the delivery: ask Gemini for bullet points of things not on the slides, so the talk adds to the deck instead of reading it ▸ 28:47…