My Playbook for Coach Bob: How to Choose the Right AI Tool

When Bob needed to call plays, I learned that sometimes the best AI isn’t the flashiest — it’s the one that delivers consistent results.

From Avatar to Coach

In the first article, I shared how Bob began as a GenAI-driven avatar with personality and flair. But avatars only get you so far — at some point, Bob needed to do something.

That “something” turned into football play calling. And it quickly taught me that the tools that make Bob conversational aren’t always the right ones when you need predictable, repeatable decisions.

Why Determinism Matters

Generative AI is fantastic when you want creativity, variety, or open-ended conversation. That’s why Bob’s avatar persona came alive through prompts, embeddings, and even a bit of emotion modulation.

But football coaches don’t act like chatbots. In a 4th-and-short situation, they don’t suddenly decide to launch a Hail Mary. Coaches — like consumers — follow patterns. Their choices may have variety, but they’re grounded in predictability.

That’s why a classic multilayer perceptron — a straightforward neural net — turned out to be the better tool. It gave me consistency without drifting off course, and that’s exactly what I needed.

Lessons from Coach Bob

A few key takeaways stood out as I built the play caller:

• Inputs Matter. My first attempt at structuring the play data didn’t work. I had to rethink what to include, how to code it, and how to frame the problem.  For example, how the game clock data was revised to emphasize its game time importance. Once I did, situational playcalling performance improved dramatically.
• Rules Still Matter. I built in constraints like “no punts except on 4th down” or “no field goals outside the opponent’s territory.” Models need boundaries just like coaches do — they can’t attempt what isn’t realistic.
• Controlled Variety Beats Chaos. To avoid robotic predictability, I added a little randomness to the system. That small tweak kept it lifelike without letting it spiral into chaos.
• GenAI is not always the best choice. Could a large LLM handle this? Yes — but only by feeding in the entire dataset for every call, which is costly and inefficient. Smaller LLMs risk drifting “off the rails.” For this job, the simpler neural net was the right choice.

Why This Matters Beyond Football

The bigger lesson here is that not every problem calls for the newest or flashiest AI. Sometimes the best method is the one that delivers consistent outcomes with less overhead.

In research, we see the same thing: some problems need generative creativity, but many rely on predictability and structure. The art is in knowing when to use which.

Looking Ahead

Of course, a coach can’t do much without players. That’s where the next step in Bob’s playbook comes in: building a team. And that’s where synthetic data enters the picture. Stay tuned for the next installment to learn how I integrated synthetic data into Coach Bob’s playbook.