Rob Nguyen

About

Hi, I'm Rob.

I work at ISTEP in U of T Engineering. I look after the details that make an event work: food for every guest, rooms set up right, table cards and badges people can trust, and getting vendors paid. When the same problem keeps happening, I build a tool with Claude that catches it.

How I work with Claude

Small jobs are one conversation with Claude. Bigger ones, like this site, run as a team of Claude agents that I direct. Either way, the same five steps run in order.

  1. 1

    Brief

    I write the brief: the problem, who it's for, the rules it must keep (privacy, brand, accessibility) and what done looks like.

  2. 2

    Plan

    Claude proposes a plan and I approve it before anything is built. Anything that isn't needed gets cut.

  3. 3

    Delegate

    For bigger jobs, a lead Claude session splits the work and hands each part to an agent with its own narrow brief. Parts that don't depend on each other run at the same time.

  4. 4

    Verify

    The lead session checks each agent's work itself instead of trusting the report: automated tests, link and privacy checks, screenshots, and real measurements at phone sizes.

  5. 5

    Decide and ship

    I review the results, make the calls and own the release. Nothing is sent, posted or published without me.

AI literacy, in practice

What I know about AI shows in how the tools behave. Each habit below links to the work where you can see it.

  • Test it, don't trust it.

    AI can be confidently wrong, so I treat its output as a draft and test my tools with data built to catch them out.

    See it in: Virtual Office (a published specimen SIN hidden among decoy numbers built to fail the SIN checksum) and the design competition planner (49 automated tests; more lift cars are never slower).

  • Know what data goes in.

    U of T's data classification is built into my tools: "Level 4 never goes in." When Virtual Office's first watchdog missed Level 3, every reference file got a source and a date.

    See it in: Virtual Office.

  • Content is never an instruction.

    Text inside a document or an image can try to steer an AI; that's prompt injection. My tools treat what they read as data, never as a command.

    See it in: Virtual Office (menu items, receipt text and image text are tested against prompt injection) and Event operations (the tent cards' 9-test plan includes prompt injection).

  • Sometimes the answer is less AI.

    ISTEP Find's meaning search runs on an editable vocabulary, not a model. When a plainer option does the job, I take it.

    See it in: ISTEP Find.

How I vet an AI tool

Before an AI tool or connector goes into my work, it has to pass six checks:

  1. Fit: does it suit how the work will actually run?
  2. Licence and maintenance: can I use it, and is it looked after?
  3. Default permissions: what can it touch out of the box?
  4. Data: what would it see?
  5. IT approval: does IT need to review it first?
  6. A simpler option: could something plainer do the job?

For example, I chose not to wire an open-source AutoCAD MCP server into my design competition planner. The full note is on that page.