Install and discover the right skills

What you'll do

Review an installation plan and identify the correct discovery route for your agent harness. Separate installed local authoring skills from Hosted read-only MCP access. You need F1-F2 and approved repository access. An agent account is optional for this synthetic planning exercise.

Establish the installation boundary

The free local core includes repository/plugin skills and runtimes under MIT licensing. That does not imply a public source repository or a generally available download. Fabric capacity and agent-provider access can have separate costs. Start with an approved checkout and its current installation guide rather than an invented public package command.

A harness is the application that discovers instructions and invokes tools. Claude Code's plugin route and a deterministic harness adapter are different discovery mechanisms. Do not assume that copying the same files into every harness gives equivalent discovery, capabilities or permissions.

Plan before applying

The brief documents these read-only planning commands from the approved library checkout:

❯node tools/agent-skills.mjs validate suites/analytics-spine.json❯node tools/agent-skills.mjs plan suites/analytics-spine.json --target <target-repository> --profile full --agents openai-codex

The target argument is a placeholder. The second command deliberately names the documented Codex adapter; it is not a Claude plugin installation command. Use the current guide for the selected harness rather than changing an adapter name by guesswork. Read the proposed writes, required parameters and target conflicts before a separately authorised apply operation.

For Claude Code, inspect the namespace and skill inventory supplied by the approved plugin route. Invoke the documented installed entry explicitly. If discovery fails, report the missing plugin, adapter or capability; do not create a lookalike command and treat it as the official workflow.

Discover a bounded capability

Choose one task: explain a source map, draft a dimensional design or review a report page. Locate the corresponding installed skill and read its prerequisites, allowed effects and expected evidence. The skill's existence does not prove that its runtime is installed correctly or that the target service is ready.

Discovery worksheet
Harness: selected from the approved installation route
Capability: one named modelling or review task
Evidence: current installed instructions and plan
Stopping point: draft or explanation before outward effect

Keep the worksheet synthetic. Do not include internal origins, credentials or client evidence in an academy submission. If a prerequisite is unavailable, complete the local reading exercise and record the missing capability honestly.

Check the resulting understanding

Ask the agent to identify the selected workflow, its inputs and its next permission boundary. A good explanation distinguishes local file generation from source/service execution and identifies which human decisions remain pending. It should not claim a clean installation or live qualification merely because a planning command completed.

Your expected output is a reviewed installation/discovery worksheet, not an applied target change. Hosted connection is a separate protected-preview workflow with Entra configuration and current project grants; a packaged connector is not evidence of a completed live login journey. Completion of this lesson provides learning evidence, not a professional credential or production access.

Check your understanding

1. Why does the Codex planning example not install a Claude plugin?
2. Which evidence should you inspect when a skill does not appear in the harness?
3. Which permissions remain separate after a successful local installation?