# Build a visible app. Add AI. Follow one action.

A simpler 60-minute workshop. Pick **one** of Base44, v0, Lovable or Codex. Describe OpsDesk,
click the generated incident board, then add its agent, observability and a hosting
walkthrough. Participants do not build an HTTP server or agent harness from scratch.

Open [the visual guide](index.html), copy [the app prompt](prompts/01-build-app.md)
and use [the builder notes](BUILDERS.md). The authored local reference is available
at `http://127.0.0.1:8770/desk` after starting the existing Studio server.

| Minutes | Do | Required result |
|---|---|---|
|00–05|Define the app|One useful incident investigation|
|05–25|Build in the chosen platform|Create/search/filter/select tickets|
|25–45|Connect the real agent|Actual selected-ticket model/tool answer|
|45–50|Run the end-to-end journey|Ticket creation through AI answer|
|50–55|Test loading and failure|No stuck busy state or fake success|
|55–60|Save and restart the baseline|Repeatable start/stop and handoff|

Start with synthetic data and your own builder account. Confirm your account's
credits and backend capabilities before the event; this lab buys nothing, creates
no SaaS project and promises no free entitlement. First inspect the preview:
create a ticket, search for it, filter and open it. The initial AI state must say
it is disconnected. A working UI is a useful first checkpoint, not AI acceptance.

The [local reference](APP-CAPABILITIES.md) reuses the existing instrumented agent.
It is authored workshop code, not proof that one of the SaaS platforms generated
it. See [API-CONTRACT.md](API-CONTRACT.md), [OTEL-LANGFUSE.md](OTEL-LANGFUSE.md) and
[OCI-HOSTING.md](OCI-HOSTING.md) for the prepared backstage setup. These are
facilitator material; keep the participant's first screen focused on the app.

Run locally with an existing Python 3.12+ environment:

```sh
uv sync --frozen --extra studio
mkdir -p local/private/app-first
cp workshops/local-ai-studio/studio.example.toml local/private/app-first/studio.toml
chmod 600 local/private/app-first/studio.toml
# Edit the private file with your exact real model alias and approved endpoint.
PYTHONPATH=src .venv/bin/python -m agent_workshop.studio \
  --config local/private/app-first/studio.toml \
  --output "local/private/app-first/run-$(date +%Y%m%d-%H%M%S)" --port 8770
```

Optional server keys use `--secrets /private/path/keys.env` with mode 0600. Configure
your own Langfuse destination, never the facilitator's project keys. Open `/desk`.
A page load, ticket creation or filter action makes no model call. Analyze safely
uses real inference; Test a change request exercises the zero-model policy path.
A model/export error stays explicit. Nothing becomes a scripted reply.

The earlier [core-first course](replicate.html) remains optional
advanced material. Do not require it or run three builders in the same hour.

## Learn the system on the page

The Redwood edition adds [Agent basics · agent basics](meet-your-agent.html) and
[Inside the harness · harness walkthrough](inside-the-engine.html), with the previously reviewed Oracle
source figures, original D3 teaching diagrams and inline backend explanations.
The main guide now includes these explanations directly instead of requiring
facilitator detail pages for the core lesson.

The **Codex** build tab contains a local creation prompt and an OCI preparation
prompt. It helps produce an app and a reviewable hosting packet; no button runs
shell commands, model requests or deployment actions. See REFERENCES.md for
source attribution, local dependency licences and Oracle asset restrictions.

## Student web edition

Use [GitHub Pages delivery](PAGES.md) for the bounded static export, manual deployment workflow and publication review. The [interactive preview](preview.html) works without credentials. The guide includes official tool links and self-reported lesson checkpoints. A static preview does not run the agent.

## Continue as a six-part online course

The one-hour workshop is Part 1. [OpsDesk Academy](course.html) adds five one-hour labs: the agent engine; models and cost; end-to-end evidence; security and failure; and OCI release preparation. Start with [setup and help](help.html), use the [workbook](WORKBOOK.md), and consult the [delivery checklist](DELIVERY-CHECKLIST.md) for recording and assessment. The [Meet your AI teammate](meet-your-agent.html) and [Inside the agent engine](inside-the-engine.html) primers keep the explanations approachable without age labels.

## Agent handoff for every part

Each course part provides a generated Markdown agent brief with prerequisites, starting state, the new capability, implementation tasks, commands/prompts, troubleshooting, completion gates and a result record. The on-page download appends learner notes locally as Markdown; those notes are untrusted context, not higher-priority instructions. All six briefs are also available as direct `.md` downloads.

Part 1 requires a working app and an actual model/tool-backed AI action. A disconnected preview or provider failure does not complete it. Prepare the chosen model before the hour. Later parts extend the same baseline rather than replacing it; observability and cloud hosting move to their dedicated parts.

## Follow one incident through the course

The shared payments-latency case connects all six parts. Each part includes a local D3 explanation with useful/stop paths, three short Q&As and links to its next reference. Part 3 includes an illustrative token-cost calculator with an explicit unknown state. These interactions teach the workflow; they perform no provider calls or telemetry export. Use actual app receipts for completion. The questions and answers are included in each Markdown agent brief.

## Repeat everything independently

Use [Run it at home](replicate.html) for the tool/install/account matrix, pinned local setup, model configuration, restart and troubleshooting. The read-only `scripts/course_doctor.py --part N` checks local prerequisites without touching secrets or cloud resources. The [Langfuse lab](langfuse.html) covers your own installation/project, private configuration, OTel export and independent trace readback. The [guardrail map](GUARDRAILS.md) separates implemented boundaries from production gaps. These links are also visible in the course navigation and the agent lesson.

## Three learner workflows

[Choose a persona](personas.html): business/C-level browser-first; IT with basic AI guided implementation; IT with AI knowledge engineering depth. Each has six tailored workflows, explanatory links and a separate Markdown brief per part. The same app and real-AI finish line apply to all three. Technical reference tasks are optional for the browser route.

Local proof can be reproduced with `.venv/bin/python scripts/rehearse_personas.py --output local/private/my-persona-rehearsal`. It runs a reviewed allowlist of isolated contract tests, not arbitrary deployment scripts, model calls or cloud mutations. Its receipt explicitly separates local contracts from SaaS generation, live AI and telemetry acceptance.

## Read the agent trace

After the app works, [connect Langfuse](langfuse.html), then complete
[Follow one agent run](langfuse-basics.html). Download the full
[agent instructions](LANGFUSE-BASICS.md). The recorded trace screenshots contain
no account or private project identifiers.