Agents read files now — so we shipped OTF as an installable skill

Six months ago, "give your coding agent context" meant pasting a wall of conventions into the chat and praying it survived a few turns. That era is ending. Claude Code, Cursor, and Codex read skills. Vercel's new Eve framework treats an agent as a directory of files. The shape the whole ecosystem is converging on is blunt: an agent's knowledge lives in files it loads on demand.

We'd been building toward that shape the whole time — every OTF kit already ships a CLAUDE.md, a .cursorrules, and a folder of tested prompts. So we did the obvious next thing: we packaged all of it as installable Agent Skills.

npx skills add otf-kit/skills

One line. Your agent now knows how to build with OTF.

The problem with pasted context

Paste your conventions into a chat and three things go wrong. The context falls out of the window after a few prompts. It doesn't travel — new project, new agent, paste it all again. And when the model forgets, it does what models do: it regenerates a button from its training data instead of extending the one you already have. You end up reviewing the same broken dialog you reviewed last week.

A skill fixes the mechanism, not just the symptom.

What an Agent Skill actually is

A skill is a folder with a SKILL.md — a name, a description, and instructions:

---
name: otf-ui
description: Build accessible web UI with @otfdashkit/ui. Use when adding
  buttons, forms, dialogs, tables, charts, or theming a React app.
---
# instructions the agent loads only when the task matches…

Agents load skills by progressive disclosure: at startup they read only the name and description of each skill — just enough to know when it's relevant. The full instructions load only when a task actually calls for them. So you can keep dozens on hand for almost no context cost, and the agent reaches for the right one on its own.

What we shipped

The free SDK — four skills, one command. npx skills add otf-kit/skills installs otf-cross-platform-ui (one API, web + native), otf-ui (web), otf-ui-native (native), and otf-tokens (theming). They're MIT, self-contained, and built on the packages already on npm.

Every kit ships its own skill. Each OTF kit now carries a skills/otf-<kit>/SKILL.md that captures that kit's stack, conventions, and tested recipes — and routes to the prompts already in the kit, so there's one source of truth and nothing to keep in sync.

It works where you already work. The skills format is supported by Claude Code, Cursor, Codex, Windsurf, Gemini CLI, Goose, Vercel Eve — 70+ agents. Install once into whichever you use.

Why this matters more than the components

Here's the part people miss: the value of a UI kit was never only the components. Anyone can render a button. The value is the context an agent needs to extend that button without breaking the other forty — the file map, the token rules, the "this is how we add a screen here" procedure. That context is what you're actually buying when you buy a kit.

Skills make that context portable and on-demand. Your agent stops guessing and starts extending — because the instructions are right there, loaded exactly when the task needs them.

Free, MIT, no lock-in

The SDK skills are free forever, and the SDK stands entirely on its own — that's the point. There's no runtime to rent and no platform to marry. Install the skills into the agent you already pay for, and own everything they touch.

npx skills add otf-kit/skills

Full guide in the docs; browse the skills on skills.sh.