Choose your harness and profile

Choose a route from the approved overview

An agent harness provides the environment in which installed skills are discovered and invoked. It does not change the meaning of qualification. Use the approved overview to choose between a harness-specific deterministic installation and the documented Claude plugin route. Do not substitute a guessed package name, public listing or plugin command. The product repositories remain private, and an approved distribution is a prerequisite to this lab.

The planning route matters because it makes proposed writes inspectable. The Codex example below validates the suite and prepares a plan using the full profile. It preserves the exact documented selectors. A profile name alone does not grant source access, authorise deployment or certify a target's readiness. Inspect the plan against the target repository before any separately approved effect.

❯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

Read the target writes

Angle-bracket arguments are placeholders. Resolve them from the reviewed learning manifest and target evidence. Do not invent a target path because an example needs one. Read required parameters, installed payload locations and proposed writes. If an existing target already contains customised files or a retained proof, keep that state visible rather than treating the plan as permission to overwrite it.

For this offline course, applying an installation is optional. The demonstrations run from the approved local library checkout. Keep the distinction in your worksheet: inspected plan, applied installation and verified installation are different observations. A script that generated a plan has not established the other two. Use current overview and package instructions for any approved apply or verify operation.

Give the assistant a bounded task

A useful learning request asks for a draft, a plan or an explanation with evidence references and a declared stopping point. For example, ask an assistant to explain the proposed Sales grain and list the human decisions needed before compilation. That request produces something reviewable without implying database-read or Fabric-deployment consent. Keep the inputs synthetic and describe only information you are permitted to share with that provider.

Separate local authoring from Hosted read-only MCP. The protected-preview Hosted service exposes authorised status and history; it does not become a remote source-query engine because an assistant can read it. Membership, exact grants and entitlements remain distinct checks. No invented organisation identifiers or implied access should appear in a practice prompt.

Keep the result comparable

Whichever harness you select, the same engineering discipline applies: human grain ratification, independent source expectations, reviewed catalogue, retained execution observations and scoped gate evaluation. A more capable model or a different discovery route cannot replace missing evidence. Generated definitions remain artifacts until the relevant observations and comparisons are recorded.

Use the following short comparison in your worksheet. It focuses on what you can review rather than claiming equivalence between unobserved installations.

Completion check

Record the harness, profile and reviewed route you selected. Explain what the plan proposes, which effects remain unperformed and why installation grants no database or deployment authority. Draft one synthetic assistant task with a stopping point. Your result should let another engineer understand the route without assuming that the harness itself supplies business acceptance or numerical proof.