For agents: the skill and llms.txt

Two files help an AI agent use ripeat safely when it only has a shell, or when it reads the web.

The skill

A skill is a short instruction file (SKILL.md) that agent tools such as Claude Code load on demand. ripeat's skill can guide an agent through installing ripeat and connecting a supported assistant with your agreement. It also teaches the safe flow: write a profile, check it with ripeat validate, preview it with ripeat plan and ripeat assess --json, show the risk in plain words, and hand the person the ripeat install command to run themselves. The one thing it hands back is the setup confirmation: ripeat mcp setup asks "Run it now?", which needs a person at a terminal, so the agent gives the command to run in a separate terminal window. It forbids running the commands that change the machine, adding --yes, --accept-risk or --allow-admin, inventing package names or recipes, and working around a HIGH or BLOCKED rating.

The skills installer can install it from the public documentation repository. It requires Node.js and asks which agent and installation scope to use:

npx skills add ripeat/ripeat.github.io --skill ripeat

Alternatively, save the file as ~/.claude/skills/ripeat/SKILL.md (Claude Code; other agent tools have their own skills folder):

mkdir -p ~/.claude/skills/ripeat
curl -fsSL https://ripeat.github.io/skills/ripeat/SKILL.md -o ~/.claude/skills/ripeat/SKILL.md

A skill is an instruction, not a lock: an agent with a shell can ignore it. The real protections are ripeat's own confirmations (--yes does not cover a HIGH-rated plan) and the administrator policy. If your assistant supports MCP, ripeat mcp setup is safer, because its tools cannot install at all unless you switch that on.

llms.txt

https://ripeat.github.io/llms.txt is a plain-text index of these docs for agents that read the web: a one-paragraph summary, then each page with its title and a link to its Markdown source (for example https://ripeat.github.io/commands.md). It is built from the same list as the sidebar, so it never lists a page that is not published.