Case Study · AI Enablement · Domain Expert Workflow

The better AI feature was teaching the expert to build without me.

Two clinicians were experimenting with Claude for intake. I reframed the problem around safety, cost, repeatability, and ownership, then built a workflow where AI generated reviewable Webform definitions that the domain expert could iterate independently.

DrupalWebformClaudeYAMLBranching logicDynamic scoringAI enablement
The Problem

Direct AI intake was easy to imagine—and harder to operate responsibly.

The initial idea was to use Claude directly in a patient-intake style workflow. I liked the underlying goal, but a live frontier-model call in every interaction introduced questions around privacy, cost volatility, consistency, and how much value the model was actually adding.

I suggested a different approach: use AI to help build the expert-authored system, rather than making AI the uncontrolled runtime system.

Architecture Decision

Move intelligence into a reviewable build step.

I created a Drupal site using Webforms and established a workflow where Claude could generate structured YAML definitions for complex forms. Those definitions could be reviewed, iterated, and then imported into Drupal.

That choice made the behavior inspectable and kept a frontier model out of the critical path for every end-user interaction. It also reduced dependence on token pricing and external-model volatility.

Enablement

Build the first two, then make yourself less necessary.

I helped administer the first two forms and taught the clinician how to work with Claude, review the generated YAML, iterate on domain logic, and bring the results into the platform.

50+ dynamic assessments.After the initial enablement, the domain expert created and refined the rest of the assessment library without needing me to hand-author every form.
Product Experience

The forms became a journey rather than isolated questionnaires.

I styled the experience and added dynamic behavior so completed assessments could influence priorities and subsequent checkups. The goal was to make the system feel like a guided path through information rather than a stack of disconnected forms.

The platform supports branching logic, scoring, changing priorities, and remediation/information paths. It is a software and enablement case study—not a claim about clinical efficacy.

What This Demonstrates

AI enablement is leverage, not dependency.

  • Exercise judgment about where AI belongs. The most obvious runtime integration was not necessarily the best architecture.
  • Turn expert knowledge into a repeatable production workflow.
  • Teach the client to operate the system independently.
  • Use structured output. YAML created a bridge between a frontier model and deterministic application behavior.
Work With Me

Have a problem that's still mostly a question?

That's a useful place to start. I work best where discovery, architecture, hands-on building, and enablement overlap.