Joan Arbó

I design systems. I teach what I learn. I build things to test the ideas.

I work between design, technology and product. I've spent much of my career building systems for large organisations. I also teach, and I build independent products to find out what happens outside the slide deck.

I might be wrong about this Interfaces may end up being an output rather than a product.

If composition is decided on the server and the device only knows a handful of pieces, what we call «the screen» stops being something you draw and becomes something you compute. I don't know whether that's liberating or terrible for the craft. I've spent two years trying to find out.

September 2026

This page is in English. The rest of the site — including every project write-up — is in Spanish. Ir al sitio en castellano.

01

Build

Things I've built.

Not scattered bets: the same head applied to different problems. Each started as a question and each is still open somewhere.

  1. In production Work

    La Vigencia

    Inteligencia regulatoria diaria: qué ha cambiado, por qué te afecta y cuál es la fuente oficial.

    Read the full story in Spanish

  2. In production Learn · Work

    Instituto Agéntico

    Formación aplicada para desarrollar criterio y práctica en el diseño de productos y experiencias con agentes.

    Read the full story in Spanish

  3. In production Live

    LORS Club

    Inteligencia de costa para decidir dónde y cuándo, sin afiliación ni intermediación.

    Read the full story in Spanish

  4. In production Live

    HabitaLocal

    Si ese local puede ser vivienda y si compensa, antes de gastar en obra.

    Read the full story in Spanish

  5. In production Live

    Repueblo

    Si merece la pena mudarse a un pueblo cerca de Madrid, con datos y no con postal.

    Read the full story in Spanish

  6. Experiment Live

    Lo Peix

    Un proyecto de espacios y hospitalidad que conecta diseño, territorio y forma de vivir.

    Read the full story in Spanish

02

Thinking

Things I've changed my mind about.

These aren't here to look good. I argued each of them in public, and something I built or something that broke made me see it differently. If I can't date what changed it, it doesn't make the list.

  1. 01 / changed my mind

    A design system is a component library.

    I used to think

    I thought a design system was mostly components, tokens and documentation. If the pieces were good enough, adoption would follow.

    Now I think

    I think it is organisational infrastructure, and that the problem is almost never technical. When someone hits a flaw in your system they don't file an issue: they quietly stop using it.

    What changed it. I learned it by being the one who stopped using it. At Amazon it was cheaper for me to build my own components than to fight the system's. It was never a formal complaint — just a quiet decision, repeated a hundred times a day by a hundred people.

    Changed my mind · 2024 The thesis that came out of it

  2. 02 / changed my mind

    Consistency is the goal of a design system.

    I used to think

    I measured success by how similar the screens looked. Arguing whether the button radius should be 4 or 6 pixels felt like important work.

    Now I think

    I think its real job is making good decisions repeatable. Consistency is usually a consequence, not the objective. A consistent system full of mediocre decisions is consistently mediocre.

    What changed it. After years of measuring adoption by counting linked instances, and getting numbers that said nothing about whether the product was any better.

    Changed my mind · 2025

  3. 03 / changed my mind

    AI will make designers faster.

    I used to think

    I assumed the big impact would be speed: the same screens, in less time.

    Now I think

    I think speed is the least interesting consequence. What actually changes is who makes the decisions, where those decisions live and how they are governed. If the judgement isn't written down somewhere, the machine invents it at scale.

    What changed it. Building pipelines where the model wrote well and decided badly. The fix was never a better prompt — it was taking the decision out of the model and putting it in a rule.

    Changed my mind · 2026 Where I check it every morning

  4. 04 / changed my mind

    The moat of a data product is having more data.

    I used to think

    I thought the advantage was coverage: more sources, more rows, more joins.

    Now I think

    I think the advantage is time. Spanish official sources publish the present and won't let you download their history. Whoever archives dated snapshots from today holds, two years on, a series nobody can rebuild. You don't buy it — you wait for it.

    What changed it. Looking for what made HabitaLocal defensible against anyone with access to the same sources. The answer wasn't in today's data. It was in having started keeping it yesterday.

    Changed my mind · 2026 The project where it surfaced

  5. 05 / changed my mind

    A place-based project can be funded with content and affiliate links.

    I used to think

    I started Lo Peix as a blog: traffic, recommendations, commission. It looked like the cheap, scalable option.

    Now I think

    I think it was a business with no moat and existential risk. Turning it into an operator moved the ticket from cents of commission to hundreds of euros per place, and the defence stopped depending on someone else's algorithm.

    What changed it. Putting both versions in a table and comparing ticket, margin, moat and risk. Writing the comparison is what changed my mind — not thinking about it.

    Changed my mind · 2026 The project, after the turn

03

Things I'm figuring out.

Not predictions, not finished essays. The questions I have open right now, with as far as I've got with each.

  • What happens when interfaces stop being designed in advance?

    If composition is decided at runtime, how much should be composed dynamically and how much has to stay deterministic? Where do the rules live, and who reviews them when they fail?

    What I've explored so far

  • Can education produce builders instead of tool users?

    AI education keeps tilting towards tools that expire every few months. I'm more interested in what happens when someone understands the system underneath and can switch tools without starting over.

    Where I'm testing it

  • How much should a product decide for us?

    Delegation is useful. Invisible delegation is dangerous. I'm interested in products that can act for you while keeping intention, limits and explanation visible.

  • How do you declare uncertainty without sounding unsure?

    Saying «this is an estimate with low confidence» has worked better for me than rounding towards safety. But I don't know how far it holds: at some point caution stops reading as honesty and starts reading as refusing to commit.

    Where I push it

04

Things I've learned by building.

What's left when you take the project away and keep the idea. Each came from something specific and works outside where it appeared.

  • Declaring an estimate earns more credibility than rounding towards safety.

    Learned on HabitaLocal, separating what is sourced data from what is a calculation with low confidence.

    HabitaLocal

  • If a text turns nobody away, it isn't finished.

    Learned on LORS: the line that qualifies best is the one saying out loud who the product isn't for.

    LORS Club

  • When the engine is separate from the domain, the next idea costs weeks instead of months.

    Learned forking HabitaLocal into Repueblo: writing the map of what gets reused before touching code is what made the rest cheap.

    Repueblo

  • Consistency doesn't come from more reminders. It comes from removing decisions.

    Learned designing a health app for people whose heads are already full.

    The thesis, in full

  • Tone is product. «The Feast» instead of «Dinner» separates a form from a moment.

    Learned on Bodados, where the register of the invitation shifted perception more than any new feature.

    Bodados

05

Teach

I learn by teaching.

Teaching is how I check whether I actually understand something. In a classroom you can't hide behind a big word: most times I've changed my mind about something that mattered, the push came from a question I couldn't answer well.

That's why the Instituto de Diseño Agéntico isn't a tools course. Tools expire every few months; what I try to transfer is the judgement to choose them, which is the part you can't download.

Instituto de Diseño Agéntico

Joan Arbó

Who writes this

Joan Arbó

I have a recurring problem: when I see something that could work better, I find it hard to leave alone. Sometimes that becomes a design system. Sometimes a product. Sometimes a school. Sometimes an experiment that goes nowhere.

I worked on Amazon Alexa's multimodal design system and led the consolidation of Western Union's global system across forty markets. I now direct design, UX and front-end at Kutxabank, and develop this methodology in public — what I publish here is mine and doesn't describe anyone's internal work.

Shall we do something together?

Three doors, depending on what brought you here. None of them requires the others.

Learn with me

The Instituto de Diseño Agéntico: judgement and practice for designing with agents.

Visit the Instituto

Build with me

If you're defining the rules of a system or building the decision layer of a product.

[email protected]