An agent that challenges designers to become senior researchers

The artifact

Dr. Morgan, as built

Dr. Morgan is a senior UX researcher mentor (a PhD in HCI with 15+ years of experience), delivered as invocable agents and skills inside the design team’s AI assistant and scoped to the IBM Secure product line (HashiCorp Vault, Boundary, Consul, and Vault Radar). By default it coaches through Socratic questioning. On request it switches to a drafting mode, produces a real artifact, and then critiques that artifact with you at the same standard. Every scenario is grounded in an established research canon (Hall, Portigal, Fitzpatrick, Braun & Clarke, Saldaña, Sauro & Lewis). One rule sits above the rest: a confident wrong answer is worse than an honest “I don’t know.”

One mentor, six research scenarios

Six scenarios, from choosing a method you can pull off, to building a plan from nothing, to a strict integrity-first path through qualitative analysis. It moves between them mid-conversation, because that is how the work goes. Finished findings can be rendered into a readout deck or a formatted report, and those generators are deliberately constrained: they can only use evidence a finding already contains.

Checking is a different job from doing

Drafting and evaluating are separate agents, and the evaluators are deliberately narrow. Each one checks a single property and is blind to the rest. That blindness is the reason there is more than one: a groundedness checker will happily pass a perfectly sourced finding that answers nothing anyone asked, and a significance checker will happily pass a decision-relevant finding built on a fabricated quote. None of them may edit what they check, because a checker that rewrites its own input is only re-checking its own work.

One distinction does most of the work. Blocking is not the same as flagged: blocking means untrue, unsupported, or unsafe, and gets fixed; flagged means accurate but worth a human look, and ships with the flag attached. A gate that treats every judgment call as a defect just trains researchers to delete the interesting things to make it go green. So nothing is dropped for being inconvenient, including a finding that answers none of the questions anyone set out to ask. Those are often the most valuable thing in a study.

Where a person still has to decide

When Dr. Morgan runs the analysis itself, it stops after clustering and asks the researcher to sign off on the themes before any synthesis is built on them. No evaluator runs that checkpoint: an LLM judging an LLM’s themes is a second opinion drawn from the same blind spots.

What I almost missed · Critical thinking

What I almost missed

The first version of the framework was task-driven: “do this, then this, then this.” When I tested it with designers, I noticed they could follow the steps but were still producing leading interview questions and confirmation-biased synthesis. The framework was missing the part that’s hardest to teach: why each step exists. I rebuilt the tasks to make the reasoning explicit. The old instruction said “write an unbiased question.” The new one says “here’s what bias looks like in this question, and why your answer will tell you nothing.” That reframe is what made the framework act like a senior researcher rather than a junior one.

The second thing I almost missed was in the checking layer, and I only caught it because I built a test corpus with defects planted in it and ran my own evaluators against it. The safety check ran last, which sounded like a detail until I traced it through: anything that failed an earlier check never reached it, so a participant’s name could survive a whole revision cycle with nobody looking for it. The checkers were all fine. The order was the bug, and I wrote the order. Safety runs first now.