SnabbSajt · Docs
For website ownersFor developers
Site Kit SDK

Agent skills

Install and verify the versioned SnabbSajt skills for Codex and Claude Code.

Install every project-local skill in one line:

npx @snabbsajt/cli skills install --agent auto

The CLI detects Claude Code (.claude/skills) and Codex (.agents/skills, also .codex/skills), installs only into the selected project, and verifies every file against the bundled manifest. Any other agent that reads a project-local skills folder can be pointed at one of those paths. Use --global only when you deliberately want the skills in your home agent directories.

Included skills:

  • snabbsajt-getting-started: the entry point. It picks the right layer for the job (local package work, a read-only delivery token, or MCP), gets the terminal or agent connected, and routes to the skill below that does the work. Start here when you are not sure which one applies;
  • import-website: inventory and convert Next.js, HTML, or WordPress evidence;
  • build-snabbsajt-site: author a new typed package, then hand it off with snabbsajt push;
  • review-site-package: verify fidelity, safety, accessibility, assets, and report honesty before import;
  • manage-snabbsajt-site: run an existing site over MCP: read it, edit the draft, and prepare a publish the owner approves. Draft-first, and it never publishes on its own.

Not in the bundle yet: figma-to-block, which turns a Figma frame into a registered block. It exists and is written against the field vocabulary the checker enforces, but the installer above does not ship it, so an agency copies it into .claude/skills/ for now.

snabbsajt skills list --agent auto
snabbsajt skills doctor --agent auto

Each skill also carries the reference material its agent needs offline, loaded only when the task reaches it: the full CLI command surface with its credential model and failure modes, the MCP tool catalogue with the scope each tool needs and what a denial means, and, for the three that author or check a package, the complete PortableSiteV1 schema, generated from the same contract the validator reads. So an agent does not have to guess a flag, invent a tool name, infer a section type, or go and read this site mid-task.

Local modifications are never overwritten silently. An update stops and shows the changed files; --force creates a backup before replacement. The receipt, skill versions, minimum CLI version, portable format, and report contract make stale instructions fail clearly.


Didn't find the answer, or is something wrong here? Tell us.

Last updated on