A three-prompt pipeline I built that turns a job description into a role-based competency framework and exports it to Excel through Power Automate. L&D teams get a defensible competency model in minutes, where it used to take six weeks of expert workshops.
Building a competency framework from scratch is six weeks of expert workshops. Building one from a job description and a competency dictionary is six minutes, if you have a pipeline that knows what it’s doing. I built Smith as that pipeline.
What I built
Three prompts in sequence:
- Decompose: extract the actual work activities, responsibilities, and required outcomes from the JD
- Map: match each work activity to competencies from a curated competency dictionary
- Calibrate: assign proficiency levels (Awareness / Working / Practitioner / Expert) based on seniority and scope
The output drops straight into a structured Excel workbook via Power Automate, formatted to match the L&D team’s existing template, so nothing has to be retyped, reformatted, or matched by hand.
What it improved
The architectural lesson here is that most useful enterprise AI is a chain of small, well-scoped prompts rather than one big model call, with each prompt doing something narrow and verifiable. When the chain breaks, you can see exactly which step broke. When it works, you have a defensible artifact at the end, not a black-box answer.
In production use, Smith has mapped 200+ roles and matches expert-built frameworks at around 85% on first pass, with the remaining 15% adjusted by a human reviewer in minutes rather than weeks. That’s the right division of labor for this kind of work: model does the structural drafting, human does the judgment calls.