Project Background:

When our organization launched an internal enterprise LLM platform, most learning teams had access to a powerful tool but no clear path to using it well. Some designers were experimenting on their own, others were waiting for guidance, and there was no shared library of what worked. The risk wasn't that people would ignore AI. The risk was that 60+ professionals would each reinvent the wheel, with no consistency, no guardrails, and no way to share what they learned.

I advocated to my leadership why our team should own the LLM platform for Talent Development. Once we had access, I demonstrated and taught junior designers and leadership new workflows using the tool and built a learning program that included a curriculum for getting started, live onboarding sessions for new users, and a library of custom assistants designed for common instructional design tasks. I also partnered with vendors to advocate for enhancements to their systems and our internal LLM team to explore integrating our learning platforms directly with the LLM stack, so AI assistants could eventually benefit from the already established single sources of truth rather than a separate tool.

The heart of the program now is how we build assistants. I treat a system prompt like instructional design: not a one-off you tweak until it feels right, but a process that is defined, grounded, and measured. I built that process into an Assistant Prompt Starter Kit that any internal team can use. Teams bring four things, and none of them require writing code: a COSTAR brief defining who the assistant is for, what it should do, and the tone it uses; context documents that serve as the source of truth, like SoPs, transcripts, and guides; golden examples, meaning five to ten human-written ideal answers we calibrate against; and the standards are already built into the kit, including a model prompting guide and GUIDE quality criteria. The kit turns those inputs into a versioned, self-contained system prompt that is grounded in the team's own sources, structured for predictable behavior, and aligned to how professionals actually learn. Then comes the part most people skip: every assistant is independently evaluated and scored with the platform's data science team using GUIDE, my open-source evaluation framework. Quality is measured, not assumed.

Today the program supports 60+ Talent Development and L&D users with a structured, supported path into AI-enabled workflows, and the kit gives every business unit learning team a repeatable route from their knowledge to a deployed, evaluated assistant. Instead of scattered experimentation, we have shared patterns, reusable assistants, and a process that will give us tight version control for future iterations.

This work grew directly out of my graduate research on RAG-enabled custom assistants, which taught me that the difference between an AI toy and an AI tool is grounding: real content, real workflows, and real evaluation.

What was my role in this project?

• Pitched leadership on a dedicated LLM program for Talent Development rather than leaving adoption to chance

• Designed the curriculum and onboarding experience that takes users from first login to confident daily use

• Built custom assistants for common instructional design and talent development workflows, applying RAG techniques from my published research to ground responses in our actual content and standards

• Created the Assistant Prompt Starter Kit: COSTAR briefs, source-of-truth context documents, golden examples, and built-in prompting standards that produce versioned, deployable system prompts

• Established independent quality evaluation for every assistant, scored against the GUIDE framework in partnership with the platform's data science team

• Developed reusable prompt patterns so teams could get consistent results without starting from a blank page

• Created ongoing self-service performance support so the program scales without a facilitator bottleneck

• Partner with vendors and internal platform teams to integrate our learning systems with the internal LLM stack

• Coach individual LLM admins and learning teams as they roll out their own assistants