UCDOps® Agent
An AI agent that runs the entire user-centred design pipeline.
Overview
An AI agent that runs the entire user-centred design pipeline — transcripts in, working prototype out, with evidence traced at every step.
The problem
Ask a general-purpose chatbot for a problem statement and you get one plausible paragraph. You cannot tell where it came from, it doesn't know what the interview said, and the next artefact you ask for won't be consistent with it. Meanwhile the manual process is sound but slow: eleven connected artefacts, each depending on the last, each needing a human to check it. The failure mode isn't speed — it's that AI output has no paper trail, so no team will defend it in a stakeholder room.
0→1
- 01
Modelled the UCD process as a directed pipeline rather than a chat surface: four stages — Discover, Define, Develop, Deliver — containing eleven skills, from research synthesis to prototype.
- 02
Each stage takes the previous stage's approved output as its only input. Transcripts, workshop notes and conversation logs go in; a working prototype of the primary flow comes out.
- 03
Built in this order: pipeline data model, evidence-chain schema, stage runner, then UI. Designing the interface last was deliberate — once the state machine was real, the interface had almost no choices left to make.
Under the hood
Stage-gated orchestration: each skill is a bounded task with typed inputs and outputs, so a failure is contained to one stage instead of poisoning the thread. Evidence links are stored as first-class relations, not prose citations, which is what makes the traceability UI cheap to render. The prototype stage generates a real, runnable front end for the primary flow.
Features & interaction design
- 01
Pipeline, not chat window
A chat transcript is a bad container for eleven artefacts that depend on each other. The primary view is a stage list with state, not a scrollback. You can always see where you are and what's blocked.
- 02
An approval gate is a first-class UI object
Nothing advances without a human clicking through. The gate isn't a confirm dialog bolted on — it's the unit the whole interface is organised around, because that's the thing that makes the output defensible.
- 03
Every artefact cites its parent
A user story links to the hypothesis, which links to the assumption, which links to the insight, which links to the quote someone actually said. Traceability is rendered inline, not buried in a metadata panel.
- 04
Comparison as onboarding
The landing page compares the agent against generic AI chat and against the manual process on four axes — stages covered, traceability, approval gates, output. Explaining the shape of the product turned out to be most of the selling.
Outcomes
- —Eleven connected UCD skills across four stages, evidence traced to source quotes, a human approval gate at every step, and eleven learn-ucd explainer pages that double as the organic search surface. Live and free to try.
- —Hours-not-weeks path from raw research to testable primary flow
- —Defensible artefacts for stakeholders and assessment contexts