How It's Made: AI Champion Training

I promised in my LinkedIn post that I'd write up how I built the AI Champion Training. So here it is — the full story of how a quick resource turned into a research-backed, gamified training platform with interactive exercises, audio podcasts, and a self-updating content pipeline.

What I want to share isn't just the tools I used. It's the workflow that emerged. Because the interesting part isn't any single tool — it's how several AI systems, each with different strengths, fit together into something that would have taken a small team weeks to produce.

Why build it at all

I kept running into the same problem. GenAI adoption is accelerating across organizations, but the people driving it — the champions, the internal evangelists, the ones actually running workshops and coaching colleagues — don't have great training resources. Most of what's out there falls into two buckets: vendor-specific tutorials that teach you how to use a particular product, or high-level strategy decks that don't help someone who needs to facilitate a training session next Tuesday.

I wanted something I could hand to another champion and say: start here. Something that covered the practitioner fundamentals — context engineering, workflow integration, model selection — but also the trainer skills that most resources skip entirely. How do you actually facilitate AI training? How do you coach someone through their first context engineering session? How do you build a community of practice that sustains itself?

I started writing a quick one-pager. But once I began building with the current generation of AI tools, the scope of what was feasible expanded fast. What started as a simple resource grew into something much more ambitious.

Phase one: research

The research phase involved three tools, each playing a different role.

I started in ChatGPT with GPT-5.2's Thinking mode to frame up my ideas. This was pure brainstorming — getting the rough shape of what the training should cover, identifying gaps in my own thinking, and generating a structured prompt I could use for deeper research. Think of this as the architect's napkin sketch.

From there, I switched to Google NotebookLM, pasting in that ChatGPT-generated prompt to gather research across sources. Unfortunately, the results were more listicle-heavy than I'd hoped. But the exercise wasn't wasted. It solidified some ideas I'd been circling and, more importantly, challenged assumptions I'd made about what champions actually need versus what I assumed they need. You can see the research notebook here if you're curious.

I then went back to ChatGPT and toggled on Deep Research for a second opinion. The difference was significant. Where NotebookLM gave me breadth, ChatGPT Deep Research gave me a robust, well-structured report that was easier to parse and build from. This became the research backbone for the entire training course — the foundational evidence base that every module would draw from.

As I've shared previously, these tools are not interchangeable. Each has a comparative advantage. NotebookLM is excellent for synthesizing across uploaded sources and generating multimedia. ChatGPT's Deep Research mode produces more cohesive analytical output. Using them in sequence, rather than picking one, gave me significantly better source material than either would have alone.

Phase two: scaffolding with Claude Code

With research in hand, it was time to build. I used Claude Code for the entire development process, but the real efficiency came from how I configured it.

Configuration 1: The writing skill

One of my first moves was creating a custom skill that mimics the writing style of McKinsey whitepapers. I took several McKinsey PDFs from their website and asked Claude's Sonnet 4.5 to analyze the writing patterns and generate a reusable skill definition. The resulting style guide captures the way McKinsey balances analytical rigor with accessibility — evidence-based, measured, consultative without being preachy. It flows well for research-backed content, reads with academic credibility without feeling pretentious, and keeps a cohesive voice across thousands of words of training material.

I chose Sonnet 4.5 specifically for the writing. In my experience, it produces more natural, human-like prose than Opus or Haiku. Each model in the Claude family has its strengths, and for sustained content generation where voice consistency matters, Sonnet has been the most reliable.

Configuration 2: Database connectivity

I configured an MCP server connecting Claude Code to my Neon Postgres database, along with a dedicated subagent instructed to query that specific database. This might sound like a minor detail, but it was a massive time saver during scaffolding. Instead of manually copying and pasting database schemas, table definitions, and sample data into prompts, Claude Code could directly inspect what existed in the database. When I said "build a progress tracking system," it could see the existing table structures and build accordingly.

Configuration 3: Phased planning

This was the most important configuration decision. When I entered plan mode in Claude Code, I gave it the ChatGPT Deep Research output, pointed it at the Neon DB subagent for database traversal, and described the high-level outcome: an interactive, gamified training platform to upskill AI Champions responsible for driving AI adoption across the enterprise.

Then I added a critical instruction: build the plan in phases, where each phase does not require knowledge of prior phases in active working memory.

This matters because of how large language models work. Context windows are finite. When you're building something complex across many files and components, the model's earlier context starts competing with its current task. By designing phases that are self-contained — each phase has clear inputs, clear outputs, and doesn't depend on remembering implementation details from three phases ago — you can clear the context window between phases. This minimizes model drift, keeps outputs focused, and dramatically improves consistency.

It's the difference between asking someone to write a book in one sitting versus giving them a detailed outline and asking them to write one chapter at a time, each with its own brief. The chapters end up more coherent, not less.

Phase three: building the modules

With the platform scaffolded, I moved to content creation. The workflow for each module followed a consistent pattern:

1. Provide the research. I'd give Claude Code the path to the relevant McKinsey-style research material.
2. Activate the writing skill. I instructed Sonnet 4.5 to use the mckinsey-writer skill for all content generation.
3. Set high-level outcomes. Each module received a brief describing its learning objectives, following the general outline from my initial GPT-5.2 Thinking brainstorming session.
4. Generate and refine. Claude Code would produce the module content, and I'd review, adjust, and iterate.

Then I'd rinse and repeat for the remaining modules. Claude Code handled debugging, minor enhancements, and bug fixes throughout the build.

The result is two learning tracks:

The Practitioner Track covers the fundamentals across eleven modules — AI strategy, readiness assessment, responsible AI, governance, operating models, RAG systems, agentic AI, communicating AI value, context engineering, model selection, and GenAI platforms. These are the skills you want before you start helping others.

The Trainer Track has four modules focused on what most online resources miss entirely — building AI literacy programs, scaling training initiatives, establishing communities of practice, and measuring AI impact. This is the material for people who need to facilitate, not just learn.

The whole thing is gamified with an XP system. Eight levels from Novice to Architect. Interactive exercises — branching scenarios, scoring checklists, drag-and-drop prioritization, matching exercises, debate simulations, guided practice forms — because reading about AI adoption isn't the same as working through the decisions yourself.

Phase four: multimedia with NotebookLM

Once I was comfortable with the written content, I went back to Google NotebookLM and created a new notebook for each training module. I uploaded the module text and used NotebookLM to generate three types of supplementary content:

• Infographics for complex frameworks and concepts introduced in the training
• Audio Overviews — perfect for people who prefer listening during a commute
• Video Overviews for visual learners

I then embedded these elements back into the training modules with Claude Code's help. Each module now has multiple ways to engage with the material, which matters when you're trying to reach champions with different learning preferences and time constraints.

The real innovation: self-updating content

The build itself — while fast — isn't the most interesting part of this story. Plenty of people are shipping things quickly with AI tools right now.

The part I'm most excited about is what happens after the initial build.

Training content has a shelf life. In a field moving as fast as AI, that shelf life might be measured in weeks rather than months. Most training programs are built once and gradually rot. Someone has to manually review every module, check for outdated references, update frameworks, add new developments. In practice, this rarely happens with the frequency it should.

So I built a self-updating pipeline. Every month, AI agents conduct research across the landscape — new model releases, shifting best practices, emerging frameworks, updated regulatory guidance. They review each training module against this fresh research and suggest specific modifications to anything outdated or incomplete.

I'm still the human in the loop. I review every suggested change for accuracy, flow, and tone before anything gets deployed. But the heavy lifting — the research, the comparison, the drafting of updates — is automated. Claude Code does the research, suggests edits, I review and approve, and Claude deploys.

This inverts the typical maintenance burden. Instead of the training degrading by default and requiring effort to keep current, it stays current by default and only requires my judgment on what to accept. The difference is substantial. It's the difference between a garden that needs constant weeding and one with an automated irrigation system that you just need to occasionally check on.

What I learned

A few reflections from the process that might be useful if you're building something similar.

Use multiple AI tools in sequence, not in isolation. Each tool has distinct strengths. ChatGPT is strong at structured thinking and deep research. NotebookLM excels at multimedia generation and source synthesis. Claude Code is exceptional at sustained development work and maintaining consistency across large codebases. The workflow that emerged — ideate, research, build, enrich — mapped naturally to different tools.

Design for context window limitations. The phased planning approach, where each phase is self-contained, was probably the single most impactful decision in the entire build. If you're building anything substantial with AI coding tools, think carefully about how to structure work so that the model doesn't need to hold the entire project in memory at once.

Invest in configuration upfront. The one hour I spent creating the writing skill, setting up the database MCP server, and configuring the subagent paid for itself many times over during the build. These aren't one-time tricks — they're reusable infrastructure that makes every subsequent task faster.

The maintenance story matters more than the build story. Anyone can ship something once. The question is whether it stays useful six months later. Building the self-updating pipeline added complexity to the initial build, but it's what makes this a living resource rather than a static artifact.

Try it out

The AI Champion Training is available now — completely free, no sign-up required. If you're driving AI adoption in your organization, or know someone who is, I'd genuinely appreciate you checking it out and letting me know what you think.

This is still early. It's mostly untested beyond my own use and feedback from a few early users. If something's unclear, if there's a topic I should add, or if you have suggestions for improvement, I want to hear about it. I'm treating this as a living resource, and feedback is how it gets better.

---

Tools referenced in this post:

• ChatGPT — brainstorming, structured thinking, Deep Research
• Google NotebookLM — research synthesis, audio/video/infographic generation
• Claude Code — development, content generation, deployment
• Neon — serverless Postgres database