My base is creative: marketing, branding, advertising, design, communication. AI isn't the goal — it's how I work faster and build systems that reach a level the ordinary can't. I encode the craft once, let the system run, and keep a human hand on everything that has to feel human. Less hype, more proof: not slides, shipped work.
I build the systems: brand understood, rules in a CLAUDE.md, a pipeline that runs only where I point it.
See the systems →En la parte visual diseño con Claude Design y Higgsfield. Con Higgsfield puedo producir vídeos visuales, campañas publicitarias y casi cualquier cosa — el mismo método de control, aplicado a la imagen y el movimiento: la marca primero, la IA solo rellena el carril que marco.
He aprendido estas herramientas a buen nivel — pero todavía no tengo un proyecto terminado que mostrar aquí. Pronto.
First, a big information dump — sources, references, screenshots, notes. Raw and messy, no order yet.
I drop it all into a structured folder system: clear names, one idea per file — built so Claude can actually understand it.
Claude goes through the folders and converts them into clean, structured notes — explained in a way it can reuse later.
We build the Obsidian maps. Now the information is connected, colour-coded by zone and far better organised — for me and for the AI. It stays updated: a constant, regulated flow, every zone attended.
From this organised brain the AI can make any proposal and the pipeline runs end-to-end — caring for each point, each zone. That's how I use it.
Brand understood and ordered, rules in a CLAUDE.md, a pipeline that only goes where I point it — here are the exact systems I've built.
A gelato brand with seven locations can't hire a writer per city, and one central voice reads flat. So I encode the brand voice once, then a Claude pipeline drafts the month's strategy, calendar and copy — re-toned per city with its own festivities and flavour of the week. Higgsfield evokes the real origin of each ingredient. Nothing publishes until it clears a 10-point anti-detection checklist: if it reads as AI, it doesn't ship. The owner approves in minutes; it runs in ~13 h/month at ~€50/month of tooling.
An AI-native fragrance brand needs evocative product imagery at scale — without the tell-tale "AI stock" look. I built a taxonomy that maps fragrance notes to image moods, then a prompt library tuned to the brand's vocabulary. Gemini personalises and Higgsfield adds the cinematic layer, but every output obeys inviolable bottle rules: the bottle leads, and no URL or QR is ever baked in — text and logo are added later, in-brand. Outputs flow straight into the brand grid.
Investment signals are scattered across dozens of feeds with nowhere to collect, score and remember them. I built a FastAPI backend that ingests GDELT, Google News across 12 regions, RSS and CoinGecko, geolocates each event on a world map, and scores its market impact with a rules-based engine — no API key, ~0 cost. Everything persists in SQLite and surfaces in a React + Leaflet dashboard, while an Obsidian vault is the system's long-term brain: signals → analysis → decisions.
Most desktop apps have no API — so Claude can't touch them. I build the bridge. A Node MCP server (single index.js, stdio) exposes 14 tools that let Claude see and operate macOS: screenshots, files, Terminal, arbitrary AppleScript. A second bridge talks to Premiere over file-IPC (8 tools, ExtendScript) so Claude can cut and assemble 9:16 reels from AI clips. These are the "arm and eyes" that make every other system run — local-first.