You run two commands, ever. After that you talk to your agent, and it drives sitekit.
sitekit is not meant to be typed by a human. It’s an alignment layer built to be operated by an AI coding agent — the CLI surface exists for the agent, not for you. If you find yourself memorizing commands, something has gone wrong. Describe the outcome instead.

1. Install sitekit

curl -fsSL https://github.com/yevgetman/sitekit-dist/releases/latest/download/install.sh | sh
sitekit --version
Should print 1.50.0. Full detail, including troubleshooting and platform notes, on the Installation page.

2. Install the agent skill

sitekit init-skill --target claude-code
Use --target opencode for OpenCode, or --target path:<absolute-path> for any other harness. This is the step that matters. It installs sitekit’s operating manual into your agent, so any session you start already knows the commands, the conventions, the failure modes, and — importantly — the things it must never do, like re-running a build over a page you’ve already authored. That’s the last command you need to run yourself.

3. Talk to your agent

Open a fresh agent session. The skill activates on its own. From here you describe outcomes, and the agent decides which commands that requires.

Starting a site

Build me a marketing site for a product called Lantern — it’s a status page for internal tools. Use the editorial style. Show me the playbook composition before you start building.
The agent scaffolds the workspace, authors the brand content, emits the pages, and runs the checks. Asking to see the composition first is a good habit — it’s your chance to redirect the voice or style before anything is written.
Where the site lands. A new workspace is scaffolded at ~/code/<product>-website by default — not in whatever directory you happened to be in. The agent will tell you the path it created. If you want it somewhere specific, say so (“put it in ~/projects/lantern”) and the agent passes --target.
Say what the product is and who it’s for, not what files you want. “A status page for internal tools, sold to platform teams” gives the agent far more to work with than “a landing page with a hero and three features.”

Choosing a look

Show me the style packs that ship with sitekit, with a sentence on each.
Rebuild the homepage in the cognitum style and show me both so I can compare.
Take the look from https://linear.app and turn it into a style pack I can reuse.
That last one captures a real brand — palette, type, and the actual motion read off the live page. See Capturing a brand.

Changing something

The pricing page header looks broken on my phone.
The tagline is wrong everywhere — it should be “Know before they ask.”
What’s drifted since we built this?
Notice these are complaints and instructions, not commands. The agent picks the right loop: a surgical section edit, a copy-layer change that fans out across every page, or a drift scan.
Don’t ask your agent to “rebuild the page” when you want a small change. Re-running a build over an authored page overwrites the body. Describe the change itself and let the agent reach for the maintenance path — the skill teaches it this, but it’s worth knowing why the phrasing matters.

Editing the words yourself

When you’d rather just type the copy than describe it:
Open the copy dashboard.
That starts a local writing surface where you can edit any page’s words and apply them — no model in the loop, brand rules still enforced. See Editing copy.

Shipping it

Check everything and deploy to staging.
Staging looks right — ship it to production.
The agent runs the checks first; if something fails it will tell you what and fix it rather than shipping broken. Deploying needs a Cloudflare account and a one-time wrangler login.

Watching the work

You don’t have to take the agent’s word for anything. Two things are worth glancing at yourself:
Run verify and show me the output.
Run doctor — is anything not set up?
Verify is the honest signal. It’s around sixty checks covering brand rules, structure, mobile, accessibility, and images, and it’s the same gate that blocks a deploy. When it fails, each failure names what’s wrong and how to fix it.

Where to go next

Building is the first of four loops. The site’s life after launch is the other three.

The lifecycle

Maintain, experiment, evolve, and adopt existing sites.

How it works

The enforcement contract behind the verify loop.

The agent skill

What your agent actually read, and how to keep it current.

Command reference

The full surface — useful for understanding what your agent is doing.