Why I Built a College Football Poll With AI

The first in a series of lessons on becoming an effective AI user

This month I published a projection of the entire college football season. All 138 teams, ranked in August, before a single game had been played. It's called thepoll.ai, the model is open source, and once the season starts, the results will grade it in public, week after week, with no edits after the fact.

I should spell out what "it" is, because the list doubles as a preview of everything this series will cover. A real working data model, built in Python from five seasons of play-by-play data. A modern website that works just as well on your phone. Live connections to real-time college football data. An email server. And a set of automations, built in a tool called n8n, that keep the whole operation running about 95% on its own.

I built every piece of that with a Claude subscription, a computer, and an internet connection.

This article tells you why a guy who runs an execution consultancy spent his summer teaching a machine to rank football teams. It also kicks off a series I'll be writing across my usual channels: practical lessons on becoming an effective AI user. And because a series of lessons should open with one, it delivers the first, about how AI picks up an accent and how to stop it.

The Poll Problem I Couldn't Let Go

First, the thirty seconds of background you need, because I promise you don't need to follow football for any of this to land.

Some college sports, including football, crown their champion differently from the rest of the sports world. There are 138 teams and each plays only about a dozen games, so records alone can't sort them. Instead, people vote. Writers and coaches publish weekly rankings, and a selection committee, working the same human way, decides who gets to play for the national championship. Think about how success gets decided in your own work: customers buy or they don't, the numbers hit or they miss, your manager signs off or doesn't. Here, the defining outcome of the whole season rests on a weekly opinion vote.

I grew up on this sport, and 1997 is the year I saw the problem for myself. The two big polls disagreed about the best team in the country, and the season ended with two national champions. Splits like that had happened before. This one I watched. The next year the sport brought in a formula, the BCS, to take the human bias out. In my opinion it never worked, because the formula blended the human polls with some computer numbers. The bias stayed inside the math.

That formula is long gone, and the bias holds on. When voters rank a team too high too early, every team that beats them collects credit for a big win that was never really there. That credit moves next week's votes, which hand out more of it. By November the rankings are standing on their own early mistakes, and they've already decided who gets a shot at the title.

A formula was never the wrong idea. That one just wasn't better than what it replaced, and the tools to build a better one didn't exist yet. Now they sit on everybody's desk.

What I Built, in Plain English

I gave the AI three rules and drew the lines it couldn't cross. It can't cheat by looking ahead: the model never sees the results it's trying to predict. It has to be obsessed with accuracy and nothing else. And no human bias gets in: no polls, and no ratings that lean on the polls.

Inside those lines, the machine did the labor. It pulled five seasons of public data through an API, which is just one computer handing data to another: 8,117 games and more than 1.2 million plays. It wrote the Python. It ran a thousand simulated seasons to test how firm each ranking is. It even threw ideas out along the way, including a clean-looking correction that made the predictions worse.

Two boards came back. The Projection looks forward: it takes how every team actually played last season, adjusts for who came back, who left, and who's coaching, and ranks all 138 teams in August. Once published, it never moves. The Poll looks backward and starts every team at zero: it reads who played who, who won, the score, and where. It can't be impressed by a reputation or a famous name. And a few weeks into the season, once enough real results exist to read, the poll starts grading the projection in public. The forecast never gets a vote in the measurement that grades it.

Why Football Instead of My Day Job

Here's where this connects to you.

My actual work is Unleash Epic. Almost every team I talk to is feeling the pressure to do something with AI, and most of them are honestly just guessing at where to start. So I built something that takes the guessing out. We look at how each person on a team turns ideas into finished work, the conditions around them, and what actually fills their weeks. Out of that comes one clear place to put AI first, plus a check thirty days later to see whether it helped. We call it an AI Execution Plan. Building that, and helping clients build serious tools of their own, left me holding a pile of practical lessons about working with AI.

The trouble is the examples. Almost every one is dry, layered in nuance, or wrapped inside someone's confidential business. I could teach you context management through a supply-chain reconciliation workflow. You'd be asleep by paragraph three, and you'd be right to be.

So I picked college football because it's straightforward. Teams play games, someone wins, and everyone argues about who's better. You don't have to know the sport, and you definitely don't have to love it. If you do, the examples will be fun. If you don't, they'll still be clear. And underneath that simple story sits a build that's genuinely complicated, dense with data, and full of the exact concepts you'd hit building something real at work. APIs. Data models. Testing. Knowing when to trust an answer.

The Tools Arrived Ahead of the Teaching

So here's my larger claim. A website. An app. A working deck. An automation that fires when a customer replies to an email, does specific things, and produces predictably consistent output with AI in the middle. If you have a computer, an internet connection, and access to any serious AI tool, every one of those is now something you can build.

I should mention the honest part. AI moves the barrier instead of erasing it. You still bring the judgment. You still have to be able to tell whether the thing you built actually works, and that's a skill you develop on purpose. I wrote about that toll in Vibe Coding 101. The capability is real, it's earned, and it has limits you can name.

What's missing is the teaching. Gallup found that AI use at work nearly doubled in two years, and in the same study only 22% of employees said their organization had communicated a clear plan for using it. By late 2025, 45% of U.S. employees were using AI on the job. Most working people were handed the most capable tools of their careers along with a login and a shrug.

So we're going to talk about it. One practical lesson at a time, each one pulled from a real build you can go look at. If you're one of the corporate refugees building something of your own, this series is aimed straight at you, and it works just as well if you're building inside a company. I'm calling the skill being an effective AI user, on purpose, because it's bigger than writing clever prompts. It's how you think, talk, and set boundaries when you work with a machine this capable.

Let me place the series on the map. Plenty of good help already exists for the first rung of the ladder: writing better prompts, cleaning up an email, using the AI built into Excel or PowerPoint. And if you want the fundamentals taught properly, go straight to the source. Anthropic publishes free courses that hit a rare balance, accessible and surprisingly useful even when the topics get advanced. This series picks up where those leave off: you, building your own custom tools. A model, a website, an automation that runs while you sleep. The lessons here are the judgment calls along the way, the ones the tutorials rarely mention.

The Proof Is Already on Your Phone

You can watch effective AI users changing an industry right now. Non-traditional media has made a huge jump in the past couple of years. Some of the best-produced, most accurate, most actionable content on YouTube, podcasts, and social platforms now comes from operations of exactly one person. Pew Research found that 21% of U.S. adults regularly get news from news influencers on social media, and among adults under 30 it's 37%.

That polish used to require a staff: researchers, editors, producers, designers. Now one person with judgment and a good working relationship with AI covers all of it. Those creators worked out the same lessons this series covers, mostly by trial and error. And one of the earliest lessons they hit explains why some of them sound like a trusted friend and some sound like a term paper.

Lesson One: AI Develops an Accent

Say you're building around a dense topic. College football analytics. Retirement planning. AI data security. Church history, if you're a pastor writing for your congregation. Every topic gets dense once you take it seriously, and taking it seriously is the reason your audience trusts you. So you feed your AI the deep material: the methodology, the source documents, the jargon. You should. That's the substance.

Then you ask it to write something for your audience, and it comes out sounding like a college professor. You tell it to loosen up. It loosens up for a paragraph and drifts right back. You rewrite, it reverts, and no amount of "make it conversational" seems to stick.

People who take this seriously usually know part of the answer already: a voice profile. A document that teaches the AI who you're talking to and how you sound when you talk to them. Students. A congregation. Fans of one specific football program. In my case, mid-career professionals on a lunch break. You build it by handing the AI writing that sounds like you, or by literally talking it through, out loud, until the profile reads like a mirror.

Here's the failure I keep seeing, and it comes from an unexpected place: the load order.

Working tools like Claude Code boot every session from a project file, a CLAUDE.md or its equivalent, that tells the AI what this project is and what to read. Most people organize that file like an index: the important dense material up top, the reference documents next, the voice profile linked somewhere below. Which means every single session, the AI reads forty pages of methodology before it learns how you want to sound.

In my experience, that order decides the accent. A model reads in sequence, and everything it has already read shapes how it reads what comes next. Boot it on the textbook and it shows up to the conversation as the textbook. By the time your voice profile loads, the accent has already formed, and it holds on the way accents do. You'll fight it in every draft and never quite win.

The fix costs an afternoon. Write the boot file itself in your voice. The very first thing the AI reads should sound like you talking, so it learns the sound before it learns the subject. The second thing it reads is your voice profile, which makes the sound explicit: who the audience is, how you talk to them, what you never do. Then it can go as deep into the dense material as the work demands, because now it knows the difference between what it knows and how it speaks.

Try the difference yourself:

Run the Accent Test on Your Own Setup

Here's your first move this week, in four steps.

Name your audience in one sentence. Out loud, to yourself. "I'm talking to overwhelmed first-year teachers" beats "my followers." If you can't name them, the AI definitely can't.

Build the voice profile. Give the AI three pieces of writing that sound like you at your best, or record yourself explaining your topic to a friend and hand it the transcript. Ask it to draft the profile: audience, tone, sentence habits, words you'd never use. Correct it until reading it back feels like looking in a mirror.

Fix the boot order. Whatever your tool reads first, a CLAUDE.md in a coding tool, the custom instructions in a chat app, the first document in a project, write that in your voice and load the profile right behind it. The dense material comes after.

Run the test. Ask the AI to explain the densest thing in your project to your actual audience. If the answer sounds like your source material, the accent formed before the voice. Reorder and run it again.

That's the whole lesson. Small, checkable, and it changes the feel of everything your AI produces with you from that point forward.

Where This Goes Next

The poll will spend the season arguing with the sport in public at thepoll.ai, and the whole model is open source. If you think it's wrong somewhere, the site tells you how to challenge it, and you'll get one of two answers back: why I disagree, or your idea built into the model with your name on the credit.

The next lesson in this series covers the boundary that decides whether you should build a thing with AI at all. You don't have to know how to build it. You do need to know how to tell whether it's built right. If that line makes you a little uncomfortable, good. That's the one to read.

You already have the same three things I used this summer: an AI subscription, a computer, and an internet connection. The rest is lessons. That's exactly what this series is for.

FAQ

What is a voice profile for AI?

A document that teaches an AI assistant who your audience is and how you sound when you talk to them: tone, sentence habits, vocabulary, things you never say. You build one by giving the AI samples of your writing, or a transcript of yourself talking, then correcting its draft until it reads like you. Stored in your project, it shapes everything the AI writes with you.

Where should a voice profile load so the AI actually uses it?

Early. The first file your AI tool reads should be written in your voice, and the voice profile should load right behind it, before any dense reference material. A model reads in sequence, and what it reads early shapes how it reads everything after. A voice profile buried under forty pages of methodology arrives after the accent has formed.

Do I need to know how to code to build something like thepoll.ai?

You need less than you think and more than zero. AI wrote the Python, pulled the data, and ran the simulations for the poll. What you bring is judgment: a clear picture of what you want, boundaries for the AI to work inside, and enough understanding to tell whether the result is built right. That last skill is learnable, and it's the subject of the next lesson in this series.