Documentation

Compound Design is Danilo do Amaral's own framework for working with artificial intelligence in design and programming. Framework, here, means a way of working with a name, rules and tools that anyone can adopt.

It is AI-native: made for people and for AI agents to use with the same care. Every part exists in two forms, one to read on screen and one for an assistant to consume, and the framework itself was built with the method it teaches.

It is not a list of tools. It is a method (plan, do, review, compound) and what it produces: data-backed articles, a catalogue of skills (instructions that teach an AI assistant to do something well) and two ways to connect that catalogue to your assistant.

The name is a thesis, not just an analogy: each thing learned should make the next one easier, not harder. Like compound interest. The next section explains how that works in practice.

Most work gets harder over time. Each new thing that comes in negotiates with the old ones, and after a few years more time goes into fighting what exists than building on it. That holds for code, for design and for the way one works with AI itself.

Compounding is flipping that. Instead of each new thing adding complexity, it teaches the system something. A fixed bug eliminates a whole category of future bugs. A pattern that worked becomes a tool. Over time the whole becomes easier to understand, to change and to trust.

The loop has four steps, and it holds whether you are fixing a detail in five minutes or building something over a week:

For people who design or build interfaces and want to use AI without lowering their own bar. You do not need to know how to code to read the articles or to understand the catalogue.

If you have never used an AI assistant to write code, start with "Concepts, no jargon". It defines agent, skill, CLI and MCP in one sentence each. After that, the rest of the site makes sense.

Three things to do right now, in order. None requires installing anything.

An AI agent is a program that uses a language model (such as Claude or GPT) to carry out tasks on your computer, not just answer questions. It reads your files, writes code, runs commands and fixes what it got wrong.

Claude Code, Cursor and Codex are agents. When this site says "your agent", that is what it means: the assistant you use to build interfaces.

An agent is only as good as the context it receives. Without instructions it produces the generic. With the right instructions it produces something up to your standard. Skills are those instructions.

A skill is a text file, written by a person, that teaches the agent to do one thing well. For example: how to animate an interface without overdoing it, how to choose typography, how to review accessibility.

It is not a program. It installs nothing on your computer. It is knowledge in text form, in the format the agent understands, and the agent follows those instructions while it works.

Each skill in this catalogue belongs to whoever wrote it. The author's name, the licence and the original address appear next to the text, always.

The catalogue gathers design-engineering skills written by dozens of authors in one place, with search and filter by topic. The list is a leaderboard: skills appear ranked by the stars their source repository has on GitHub, a public signal of how many people found it useful.

Some skills open right here, with the full text. Others open at the source site. The difference is the licence: we only serve the text of those who allowed redistribution. The rest show the card and take you to the original.

CLI stands for command-line interface: a tool you use by typing commands in the terminal, the text window of your computer, instead of clicking buttons.

The Compound Design CLI is for fetching skills without opening the site. You type a command, it answers with the list or with the text of a skill. It is mostly useful for the agent, which lives in the terminal, but it works for people too.

No install needed. The command below downloads and runs on the spot.

MCP stands for Model Context Protocol: a standard way for an AI agent to talk to an external source of information as if it were part of the agent.

Think of a power socket. Before MCP, each agent needed a different adapter for each source. With MCP, any agent that speaks the protocol connects to any source that speaks it too.

Connecting the Compound Design MCP to your agent means it can see the whole catalogue: list topics, search skills and read the text of one, without you copying anything by hand.

Claude Code is Anthropic's agent that runs in the terminal. It speaks MCP. One command connects the two; after that, the catalogue shows up to it as a set of tools.

If your agent does not speak MCP, or if you prefer to see what you are handing it first, use the CLI. The command below prints the text of a skill in the terminal; copy it and paste it into the conversation with the agent.

Replace emilkowalski/animate with the name of any skill in the catalogue. The name is at the top of its page and in the copy button.

Each article exists in two modes. Reading, as a normal text. Presenting, where the same blocks become slides, with the charts full screen, for you to use in a meeting.

The text and the slides are the same file. If a number changes in the article, it changes in the slide. Nothing is typed twice.

Each chart in the articles reads its data straight from the survey. No value is typed by hand. When an article cites "82.6%", that number came from the same file that drew the bar beside it.

An automatic check compares every percentage cited in the texts against the dataset, and the site does not publish if any diverges.

Each skill in the catalogue belongs to whoever wrote it. The licence of each source repository was checked one by one. The permissively licensed ones, the majority, are served from here, always with author, licence and the exact commit they came from, so you know you are reading the same version the author published.

The ones that declare no licence appear in the catalogue but open at the source. Without a declared licence all rights stay with the author, and serving the text from here would mean redistributing what is not ours.

This site practises what it describes. Each fix that repeated became a check that runs before publishing and breaks the build if anything diverges: the numbers cited in the articles, the skill links, the counts written in the prose. Each architecture decision has a record in docs/adr. It is the compounding step, applied to the product itself.

The leaderboard ranks by the source repository's stars on GitHub. A star is how someone marks a project as useful; it is a public signal anyone can check.

We do not measure how many times a skill was installed. Nothing is installed through here, so that number does not exist, and inventing it would be the opposite of what this catalogue stands for. A heavily starred repository appears at most twice in a row; the rest is collapsed, or a single author would take the whole top.

The article series uses State of Prototyping Spring 2026, a UX Tools survey of 1,478 designers and builders, collected between March 14 and April 6, 2026 across 18 regions, under CC BY 4.0.

The aggregate data is versioned in the repository. A script converts it to the format the charts read and computes the derived values, such as the three vibe-coding groups. Today that is 204 numbers verified, with zero divergences.

Comment on the presentations, propose a skill for the catalogue or an article. Content with a draft label is exactly what is waiting for review.

Content lives in text files in the repository: articles in MDX, the catalogue and the index in JSON. Editing does not require touching code.

The survey data belongs to UX Tools, under CC BY 4.0. The texts on this site are by Danilo do Amaral and may be cited with credit. Each skill in the catalogue carries its own author's licence, shown next to the text.

The CLI code is MIT. The site code is open in the repository.