---
title: "design-system.ai"
url: /
description: "Design systems that AI agents can read and obey. The method from five design systems over fifteen years, with the rules, the checks and the gaps in public."
---

## §0 · thesis

# Design systems have no rule book that agents can obey.

An agent has no tacit knowledge. It reads what is written and obeys what is enforced. No framework says which rule wins, how strict each rule is, or what proves that a rule holds.

*Pavol Hudran. I have built design systems since 2011. This site is what I learned, with the evidence.*

[Read the white paper](/living-design-system) [Find your level](/ladder) [Slides, TechMeetup 2026](/talk/techmeetup-2026)

## §1 · vision

### Where this is heading.

A design system is becoming the contract an agent works under. The parts already exist: a rules file, tokens with a resolver, a component registry and an open protocol for agents. What is still missing is a shared rule book: which rule wins and which check proves it. [Read the white paper](/living-design-system)


[Read the vision](/vision)

## §2 · migration

### Fifteen years, five of my design systems: where the decision was buried, and where it moved.

Each time, I did one thing: I took a decision out of where it was buried and put it where the next reader could use it. In 2026, the next reader is an agent.


| Era | Buried in | Moved to | Read by |
|---|---|---|---|
| 2011 · Makro, M-Objednávka | one project's skin | a method, several build targets | several teams |
| 2015 to 2016 · ČP Invest, Generali Investments | a CSS framework's compiled output | variables and a skin layer | one team |
| 2022 to 2024 · Notino | design files and designers' heads | tokens as data, versioned packages | 15 internal teams [claim case3-teams: The author's reference deck, 2024; source: author's reference deck 2024 (docs/early-work/phudran_reference_2024.pdf); confirmed] |
| 2026 · rules for agents | people's tacit knowledge | numbered decisions, each with a named check | AI agents |

MCP tells an agent what exists. It does not say how the team uses it, which rule wins a conflict, or whether the result is correct. That layer is open. [The white paper](/living-design-system) describes what it needs.

## §3 · same code


### Only the reader changed.

This site shows the same move, live.

One token set. One resolver with two modifiers, `brand` and `mode`. Four permutations, no component changes. Drag the seam, or add `?brand=kiln` to any address.

## §4 · ladder

### Five levels. Each answers one question about a decision.


Levels 1 and 2 are not obsolete. Weak tokens give an agent nothing solid to select from. The step from 3 to 4 is not a purchase. It is a list of decisions, an order of authority and checks.

[Start the self-assessment](/ladder)

## §5 · start

### Four steps to start this week.

For you, if you run a design system and want coding agents to follow it. These are not prompts. They are four changes to your own repository, small enough to start this week.

1. Spend one hour counting what your system has against the [Design System Checklist](https://www.designsystemchecklist.com/).
2. Write a short rules file that names what the agent gets wrong, with the correct replacement.
3. Number your decisions and add an "Enforced by" column. Leave it honestly empty where nothing checks the rule.
4. Add one check that runs in CI for the rule that hurts most.

## §6 · proof

### This site is built by the method it describes. Check it.

- **Same code, two brands.** Press the switch in the header. Components do not change; only `data-theme` does.
- **The register, in public.** Every rule names its check, or says `not enforced (gap)`. [All decisions](/decisions)
- **Every number is on record.** Each figure on this site is an entry in a public registry, with its base, source and date. [The claims registry](/claims.json)

## §7 · open source

### A brand voice guide your tooling can prove.

Voice and Tone is my open-source plugin for Claude. It turns a vague voice guide ("warm, but professional") into rules an agent can follow and a reviewer can check. Every rule records where it came from and how sure it is. Tone is computed from 80 tone cells from 18 stored rows [claim plugin-cells: The plugin's public pages, read 2026-11-16; source: https://pavolhudran.github.io/ai-voice-and-tone/; confirmed].

It is built on Mailchimp's Voice and Tone framework, not on a design system. What it shares with this site is the method: evidence before authority, and checks in code.

[What the plugin does](/voice-and-tone)
