You do not need everything on day one.

1. Get the reviewed workshop revision from the published repository. Open a terminal at the repository root.
2. Check Python 3.12+ and uv are installed. Run uv sync --frozen --extra studio.
3. For an API-only or locally hosted model, obtain the exact alias and confirm native streaming/tool support. A coding-assistant subscription is separate from runtime API entitlement.
4. Copy the example TOML from workshops/local-ai-studio/studio.example.toml into an ignored private directory. Set file mode 0600. Edit only your own endpoint/alias and environment references.
5. Start with telemetry disabled. Prove one real tool-call task, then enable your own Langfuse destination. Never copy the trainer’s credentials.
6. Use a fresh output directory for each server launch. If port 8770 is occupied, identify its owner and choose another port; do not kill an unknown process.

Mac, Linux and Windows with WSL can follow the terminal commands. Browser-only participants can complete the product sketch and architecture exercises but cannot claim runtime acceptance.

Keep source changes, the exact command, synthetic input, actual outcome and an evidence reference. Raw answers, traces, recordings, endpoints and secrets stay private by default. Use the workbook for a sanitized summary; task checkboxes do not prove execution.

The local Studio supports real streaming model/tool calls, session-scoped JSON ticket actions, bounded registered synthetic tools and explicit Langfuse OTel export. Its model and backend acceptance still depend on your configuration. The static product sketch has no AI. The hosted OCI coordinator and its telemetry integration use a separate protocol/configuration; Part 6 makes that adapter work explicit.

The course pages and local tests are supplied. Cloud accounts, model credits, recorded course videos, production authentication and a running learner sandbox are not bundled.

One operator, one ticket inbox and one read-only investigation. You do not need a database or cloud account to explore the product sketch.

Try asking: “Which information would let an operator decide their next safe check?” Write that down before choosing a model.

Create a new synthetic ticket, search for its title, filter it out, clear the filter and select it again. A screenshot alone cannot prove these controls work.

If a control is decorative, ask the builder to implement only that interaction and retest it. Keep the AI panel disconnected until the controls work.

Check whether the request reaches your own server, then inspect its sanitized error. Confirm the exact configured model alias, authentication binding and native tool support.

Start with one ordinary greeting, then one health question. Change one setting at a time; never substitute a canned answer for a failed model call.

Separate local recording, exporter delivery and backend readback. Check the exporter destination and auth without printing keys; then search the exact trace ID in your own project.

A missing trace is a telemetry failure, not proof that the model failed. Do not repeat paid inference just to repair a collector.

Approval is an application control. The change intent is stopped before model construction, even if a prompt asks it to ignore policy.

Run the policy exercise and inspect its zero-generation result. Then compare an allowed read-only question with an unavailable write-tool request.

No. You can finish the local app and explain its target architecture. A live hosted invocation needs your own scoped tenancy, identity, image and network path.

Use the deployment-preparation prompt to produce a reviewable packet. Keep a source walkthrough labelled as such until a hosted invocation and its trace are verified.