In the age of AI, a core responsibility of the L&D leader is to understand how large language models (LLMs) such as ChatGPT, Gemini, or Claude interact with our people. One important caveat is that, when used out of the box, these LLMs provide generic output that doesn’t apply to role- and organization-specific situations. 

‍Context engineering can help transform LLM output into relevant, targeted, and adaptive content learners can access within their flow of work. 

Context engineering acts as an intermediary layer between external LLMs and our organization’s people and work functions, serving two vital purposes:

• It draws upon a rich repository (see “Three Foundational Steps Toward AI-Powered Adaptive Learning”) of institutional wisdom, business practices, processes, policies, and procedures to provide relevant, targeted responses to learner queries, and 
• It protects our organization’s data and intellectual property from becoming part of an LLM’s public knowledge base—and potentially passed along to other users. 

Context engineering also helps us create truly adaptive learning experiences without the long runways of the traditional learning design process. See below for some of the robust functions AI can enable within our learning solutions.  

Three Foundational Steps Toward AI-Powered Adaptive Learning

If you’re feeling inspired by these possibilities, you may be wondering about the development process. Here’s a high-level summary of the three foundational steps that lead to an AI-powered learning solution: 

1. Assemble a knowledge base: LXDs and stakeholder SMEs work together to assemble a robust repository, or knowledge base, of information relevant to the learning objectives. This information might include manuals, standard operating procedures, documentation, videos, images, slide presentations—as well as anecdotes, hacks, and wisdom from seasoned in-house experts. LXDs and SMEs must then organize and curate these materials to ensure that the knowledge base contains consistent, reliable information.

2. Produce a RAG: After LXDs and stakeholder SMEs devise a learning experience that meets the objectives, LXDs work with developers to produce a retrieval-augmented generation (RAG) database. A RAG extracts the information from the knowledge base that’s needed to support a particular learning experience. Many RAGs can be constructed from a single knowledge base for different learning needs. 

3. Develop prompts: LXDs then develop prompts that tell the AI about a learner’s role, what it’s supposed to do, how it should behave, how it should teach, and what feedback it should give. AI might serve as an evaluator, coach, assimilator, curator, simulator, or a combination of two or more roles (see left/right/above/below).

For example, a prompt might begin as follows: 

You are a friendly workplace coach helping learners practice {skill}. Use ONLY the information in the provided documents to provide feedback on the following statement: {statement}.

Creating a prompt is not a one-off task; it requires continual evaluation, validation, and iteration. Our team must constantly review what the LLM is doing, evaluate whether content is fulfilling the learning objective, analyze how it’s presented to the learner, and refine the prompt accordingly. 

Future Frontiers in Context Engineering

As AI tools evolve, we’re likely to see opportunities for knowledge tracing, or the process a machine uses to track a learner’s knowledge as they interact with a learning experience. Knowledge tracing helps us create truly adaptive learning: As our AI tools uncover which content is most effective and how learners interact with it, it can then automatically adapt, update, and evolve in response to learners’ reactions.

Context engineering is already offering us fresh ways to extend the reach of human expertise. For example, until recently, a seasoned expert’s collection of insights and “war stories” often reached only a few team members and friends. Context engineering can help us capture this heretofore intangible expertise within an internal knowledge base and make it accessible to a wider audience. 

Context engineering can also help us extend high-touch, high-value learning and skilling experiences to a wider audience. One-on-one coaching is a great example: Experts have limited time and are in high demand, so it’s difficult for them to offer extensive coaching to new or junior team members. However, an L&D team might partner with an expert to develop an AI-powered experience that helps learners practice a targeted skill. 

To get the process started, we recommend working with specialized vendor-partners to develop a context engineering strategy for their people. (Caveat: Before inviting a prospective partner into the conversation, it’s important to verify that they are fully committed to protecting your data and intellectual property, and have the credentials—such as ISO 27000 or SOC 2—to prove it.)

The innovative learning team at Hilton partnered with SweetRush to do just that, via an immersive AI coaching experience that helps learners practice resolving a guest problem using a five-step model. After many tests, refinements, and iterations, the Hilton-SweetRush team trained a large language model (LLM) to recognize examples and nonexamples of the five steps, then deliver feedback based upon the learner’s performance. 

Learners face realistic digital avatars and speak aloud to practice the five-step model, with a choice of 20+ scenarios and unlimited opportunities to try again when the situation doesn’t go as planned. It’s an ideal use case for AI: helping humans see, hear, and understand each other better. 

If you’re looking for a way to extend the reach of your experts and scale high-value learning and skilling experiences to a wider learner audience, we’d love to explore the possibilities. Reach out to share your vision with one of our (unpromptable) experts.

‍