Skilling up: How to use an LLM for instant, custom training
When it comes to workforce impact, I am mostly worried about replacement and deskilling, but I thought I’d take a walk on the more optimistic side :) How can LLMs help you improve your career outcomes?
In this post, we will talk about how to use LLMs to develop instant, customized training for a niche skill you are trying to learn. I’ll use an example from my work (as a researcher at the Policy and Innovation Center) to show how Large Language Models can streamline the process of learning how to do something new.
The Problem
Examples of Sankey diagrams
Say one of my analysts needed to create a Sankey diagram: a specialized visualization that shows flows between categories. It’s not useful for too many things, and we’ve all made it this far without needing to know how to build one, but it is perfect for tracking how workers move between occupations.
She would face the typical journey of teaching yourself a technical skill online:
First, she'd need multiple Google searches to answer basic questions:
When are Sankey diagrams (not) appropriate?
Can our usual software create them?
Which code libraries can make one? Among those libraries, which makes the best one?
What are the dimensions of “best?” Clear? Pretty? Customizable? Easy to make? Compatible with other data visualization libraries we use?
(Once she’s identified the one she wants to use) How does she use it?
Each question spawns more questions. She'd spend time comparing solutions, weighing trade-offs, and piecing together scattered tutorials. She’d rely on forum posts as sources often, which requires adapting questions slightly different from hers to her situation and evaluating the credibility of answers.
This process teaches you a lot of extra information about Sankey diagrams, which is valuable if Sankey diagrams will become part of your work often, or if you need to teach someone else how to build one. But if this is a one-time project? If it has a tight deadline? The process of digging through search results, forum questions, and code library documentation becomes a barrier to getting work done.
The AI Option
If she decided to use an LLM to learn this skill, she could write a prompt that looks something like this:
"I would like to learn about Sankey diagrams. First, I'd like to know what they are and are not useful for. Second, I'd like to know how to build one. I am using [our statistical coding language], and I would prefer libraries that are compatible with [software package], but I can also use Keynote, PowerPoint, or Excel if those are substantially better or easier. Please select software, list the other options you considered and explain why the one you selected is best, and explain how to use the one you selected to make a Sankey diagram."
I tried this, and the AI response included:
A clear table showing appropriate and inappropriate use cases for Sankey diagrams
A comparison of software options with compatibility notes and trade-offs
A specific recommendation with justifications
Step-by-step instructions with working code and explanations
Guidance for common next steps like web deployment or presentation integration
Quick Tip: If this doesn’t work well for your case, try reducing the LLM’s task complexity by splitting it up into different prompts in the same chat. For example, first ask, “what is [skill] and when is it useful?” Then ask, “What are the types of . . . ?” then, give it some context and ask how it might apply in your situation. Once you’ve walked it through all the background you think you need, you can request your table and recommendation.
When AI Training Hits Reality
When I tested this, there was a pretty big hitch: the code it gave me did not work!
But here's where AI training shows its strength: I could immediately ask for a fix. The AI explained what went wrong, provided corrected code, and outlined why the changes solved the problem.
This interaction reveals something important about AI-powered learning. It is not perfect, and, just like any other output from LLMs, you need to test what it says. However, if you are prepared for that, it can be a responsive learning partner that can adapt, correct, and clarify in real time.
Four Principles for AI-Powered Training
1. Write Strategic Prompts Generic questions like "How do I make a Sankey diagram?" produce generic answers. Effective prompts include your constraints, preferences, and specific needs. Consider starting even broader: "I have data with [attributes] and want to visualize [pattern]—what visualization options should I consider?"
2. Maintain Critical Thinking Test everything. Request sources when learning concepts, not just procedures: "Please cite sources for each fact claim, preferring peer-reviewed sources when possible. Note any uncertainty in brackets." This approach tends to surface better quality sources and make it easier to fact check.
3. Engage in Dialogue Don't treat AI like a search engine. Challenge responses, ask follow-ups, and request corrections. The ability to iterate makes AI training more dynamic than traditional resources.
4. Know When to Start Fresh If conversations drift off-course or become less helpful, start a new chat. You may have reached the end of your context window, or it may have collected some mistaken impression about what you want.
When to use LLMs for skill development
This approach doesn't replace deep learning when you're building core competencies. But for those "I need to figure out how to do X quickly" moments, AI provides something new: instant, customized training that adapts to your specific context and constraints.
What learning challenges in your work could benefit from this kind of instant, tailored instruction?
—
LLM disclosure: I asked Claude Sonnet 4: “Can you write a blog post based on this chapter?” and I pased in a case study from my book manuscript :)