I have been designing for banks for over 10 years. Now I build everything from scratch with Claude, Midjourney, and PHP. This is not theory. It is what I do every day.
What this is
A trilingual technical blog about AI, GEO and UX design. I document what I build, what fails, and what I measure. No generic theory.
Who it is for
Developers, designers and SEOs who want to optimize their content for AI agents, not just Google.
The principle
No stock photos, no generic prompts. Everything built, tested, and documented with Claude and PHP in real production.
Most connected
What is llms.txt and how I implemented it on my blog
The llms.txt file is the equivalent of robots.txt for artificial intelligence. Here is how I built mine and validated i…
14 connections
Explore
My Real Workflow: 4 AIs I Use Every Day as a Designer
How I use 4 AIs every day as a UX/UI designer: Claude for code, Midjourney for images, ChatGPT for ideas and Gemini for…
11 connections
Build
How to write AI prompts: a designer's guide (2026)
What is a prompt, how to write prompts for Midjourney and ChatGPT, and why designers have an advantage writing AI promp…
10 connections
GPT-Vision analyzed a screenshot of my homepage and then the HTML of the same site. From the image: 'I do not know whose site this is.' From the HTML: 'very strong semantic structure.' The same site is more legible to machines than to humans.
Two of the three AI models I use daily insert invisible watermarks in the text they generate. The third does not. None of them distinguishes between text the AI wrote alone and text a human dictated and the AI transcribed. The Ship of Theseus has a new problem.
After 6 months building with Claude, I learned that delegating to AI is not giving it approval. It is creating a review step that did not exist before. There are things I always delegate. There are things I never delegate. And the line moved more than I expected.
This blog publishes in Spanish, English, and Japanese. I do not translate from one to another. Each language has its own version of the same topic. This is the process, what changes between languages, and why it matters for AI visibility.
The llms.txt generator was born because I felt hypocritical writing about something I had not tested. Generative identity was born because stock photos feel dead. The graph search was born because search engines feel broken. GEO Tarot was born because I wanted to play with GEO.
AI did not teach me new things. It forced me to do things I always should have done but that workplace culture normalized skipping. Asking deeper, documenting for others, justifying technical decisions, and questioning whether something works or has simply not broken yet.
Four AIs read the same llms.txt and all four correctly identified me as a UX/UI designer with fintech experience. In the previous test, using compressed memories, two of them got my profession wrong. A plain text file communicates who I am more clearly than months of conversations.
I asked Claude, ChatGPT, Gemini, and Perplexity to predict my answers to 10 questions. All four built a more rigid, more principled, and more consistent version than the real me. And that version gives worse advice.
I asked Claude, ChatGPT, Gemini, and Perplexity to tell me everything they know about me without searching past conversations. Four different profiles. Four levels of detail. None could tell me which conversation they learned each fact from.
I gave the same URL and the same prompt to GPT, Perplexity, Kimi, Meta AI, and Mistral. Four accessed the real data. One fabricated methods, IDs, and structures that do not exist. And every single one requires registration to use.