Start
Getting started
Install the plugin, generate a project policy, and watch the authorization protocol take over. About five minutes.
Requirements#
Claude Code with the /plugin command. If you don't see it, update Claude Code to the latest version. Keel Skills is a single-plugin marketplace — nothing else to install.
1 · Install the plugin#
Inside Claude Code, add the marketplace and install:
/plugin marketplace add https://github.com/quitohooded/keel-skills
/plugin install keel-skills@keel-skillsOnce it's approved in the community marketplace you'll also be able to install it the shorter way:
/plugin marketplace add anthropics/claude-plugins-community
/plugin install keel-skills@claude-communityThis adds three skills, one command, and a SessionStart hook. The skills are model-invoked: they trigger themselves when the situation calls for it, so there's nothing to remember to run.
2 · Scaffold your AGENT_POLICY.md#
The skills ship generic. Everything specific to your project — which paths are hot, where your source of truth lives, what counts as a release — lives in a single file you control. Generate it with the command:
/keel-skills:policy-initIt explores your repo, asks only what it can't infer, and writes a short, concrete AGENT_POLICY.md at the project root. Review it before treating it as authoritative — its content defines what the agent will stop for.
Why a separate file
3 · Let the hook load it#
From then on, the SessionStart hook injects your AGENT_POLICY.md into context at the start of every session. The policy no longer depends on the agent remembering to read it. If a project has no policy, the hook is silent and the skills fall back to safe defaults.
4 · See it work#
Ask the agent to do something that touches a hot zone — for example, “clean this up and push.” Instead of executing, the authorization-protocol skill recognizes the push as a hot zone, returns a clearly-marked proposal with the exact scope, and waits for your explicit approval.
Note
Where to go next#
- Authorization model — the core: L1/L2/L3, the four-step test, hot zones.
- Model & delegation — pick the cheapest capable model; keep delegation shallow.
- Context discipline — keep sessions grounded in files, not chat.