Vibe Coding, Clinical Tool Building & Agentic Workflows

The Diagnostic Promptᴵᴹᴵ Applied to System Design + Writing the PRD

The session revisits the Diagnostic Prompt, this time applied to a tool idea: before anything is specified, AI questions the problem itself to bring out hidden assumptions about the user, edge cases, failure modes, governance needs and scope creep. The output is a Product Requirements Document (PRD) with three scaffolds: clinical (who uses it, in what moment, what decision it supports, what it must never do, which populations it applies to, what evidence it rests on), technical (where it runs, where data lives, how it is hosted) and user experience (the user's journey under pressure, with three to five worked cases that define correct output). The clinical scaffold is where a tool is placed on the right side of the TGA software-as-a-medical-device boundary, and where algorithmic bias is addressed, including for Aboriginal and Torres Strait Islander patients. The facilitator writes the PRD for a reference build live. Participants start their own in the session.

Writing Your PRD with the Diagnostic Promptᴵᴹᴵ

Participants run the Diagnostic Prompt on their own tool idea and record the AI's questions word for word. They then write their PRD across the three scaffolds, including the worked cases, and set it up as a repository with the agent instruction files that point to it. No code is written this week. Participants also complete the vibe coding environment setup checklist before Week 2.

Building and Testing Locally

The session opens with the Custodian Promptᴵᴹᴵ applied to the development environment itself: what the coding agent can see, what folder is open, what leaves the laptop. It then takes the reference PRD through a vibe coding environment to a running tool on the facilitator's machine, and shows the same PRD driven through two other coding agents to demonstrate that the PRD, not the coding tool, is the asset. Testing is taught as a clinical discipline: the worked cases from the PRD are run as tests, guardrails are challenged with adversarial inputs, and the Convergence Checkᴵᴹᴵ is applied to the tool's outputs across repeated and reworded cases. Plain-language debugging and a validated ten-item usability instrument (how to administer it, score it and read the results) complete the session.

Testing Your Tool Against a Validated Standard

Participants build their tool locally from their PRD, run their worked cases and adversarial inputs as tests, and fix what fails. They then give the usability scale to at least two colleagues who weren't involved in building it, calculate the scores, and reflect on whether the tool is ready.

Deployment and Agentic Workflows

The session covers the step from a tool on a laptop to a tool at a public URL: the Custodian Promptᴵᴹᴵ run again on the deployment, what changes when a stranger can find the tool, and what professional indemnity does and does not cover for a clinician who publishes software. It then adds an agentic layer to the deployed tool: the difference between a triggered automation and a scheduled autonomous agent, and the coding agent building its own schedule so the tool is reviewed on a timer without manual initiation. Mandatory human-approval checkpoints and a pre-deployment safety checklist are covered. The reference build is deployed live.

Deploying Your Tool and Measuring the Agent

Participants deploy their tool to a public URL and add one scheduled agent that runs against synthetic or de-identified data. They run it at least three times, document its accuracy, the corrections it needed and the oversight it required, then reflect on what conditions would make it safe to run without review.

Walkthrough of Your Deployed Tool

A facilitated walkthrough of each participant's deployed tool. Another participant uses it as intended, without guidance from the builder. The group covers the problem it addresses, what works, what fails, and what would need to be true before a colleague could use it. Each participant leaves with a one-page governance summary assembled from their PRD, test results and agent runs. There's a prep guide for this session.

Reflection on the Full Lifecycle

Participants write up what they learned across PRD, build, test, deployment and agent, including how the PRD changed along the way and what is still needed before a colleague could rely on the tool.