Practical AI: Magic Draft Style

"But what do you DO with AI Dave?"

When I tell people I spend time in straight up Chat-bot AI every day, I often get the “Huh?” response. I get it. Asking the unreliable answerbot questions gets old after a while. And while I love dwelling in the deep end of the pool on the BIG AI QUESTIONS, I just thought I’d point out: it’s already a hell of a consumer product if your work or even better, your hobby, involves information of any kind for one simple reason: Time. ChatGPT (and Perplexity, and Claude) manufacture new time in my day, and that’s invaluable.

I currently pay $20 a month for each of the commercial AI services (because of course I do), but I’m finding I spend most of my time in the current ChatGPT interface for the 4o model (although shoutout to the new Advanced Voice feature as a research bot on a long car ride. No internet access, but for anything trained, wild!)

The Context

Strap in NERDS, it’s time to talk about the greatest game ever made:

I’m not going to explain how the game works or tell a long story: it’s a great game I’ve been playing and writing about for centuries at this point. It’s a big complex collectible card game with byzantine rules, and I’ve been playing a particular format called “Draft” since the early 1990s. 8 people sit at a (now virtual) table, each with three packs of 15 cards unopened. Everyone opens a pack at the same time, picks the one card they want, and passes the deck to the left. Do it over and over until everyone has 45 cards, make the best deck you can, and play everyone else. The “fun part” is the insanely complex interactions between all possible cards in the set and making decisions about strategy on the fly. It’s like scratching my brain with a toothbrush.

The “not-fun part” is learning the new 150-300 cards that make a new “set” a few times a year. It’s a lot of reading card text and making connections (or reading other peoples’ analysis).

If that all sounds alien, just swap all the “Card Game” stuff for “Fantasy Football” and it’s just stats and player ratings and all those podcasts and analysts who rate each year’s crop of rookies, except you draft a few times an evening, not just once a season.

The Problem

I’ve been learning the new set (Duskmourn) and having a really hard time with it — I’m clearly missing some concepts. Normally, I’d spend an evening on reddit reading peoples lists and comments on every card. This time, I decided to have ChatGPT help me.

How do I learn something faster than just reading? By having a coach while actually doing. Could I make an AI that could help me draft live? The equivalent of having a coach in my ear?

The Prompt

If I’m doing real work, I know how to write a long, narrow guiding prompt to skip all the iteration, but most of the time I’m not even sure what I want, so I started stupidly simple talking like I talk:

I then pasted in a screenshot from the excellent folks at thegathering.gg who’s system of ranking cards useful, and it gave me a wordy all clear:

I then pasted in a screengrab from the actual draft hand I was playing (against bots, so with no time pressure) on the crack-cocaine-adjacent Magic: The Gathering Arena. One reason I’m hooked on the 4o model from OpenAI is all this multi-modal, images, PDFs, audio, whatever just seems to work with cut and paste.

To my surprise, it immediately returned the following:

For each card in the hand. All that analysis is is bland but accurate and based, I assume, on training data (there was no visible call to the internet for outside help.) One assumes the CopyrightBorg scooped up countless articles on draft strategy. (Which, at one point, I used to write).

Already, this is pretty darned amazing. But I wanted something tight that let me just go through hand after hand double checking my intuition and seeing where I was over and undervaluing certain cards. So I asked:

Let me be exquisitely clear: this is not how you work with an AI. This is me just talking to it. Badly, and somewhat on purpose.

And it went and made a python program to access the database and present a tight table:

I mean. Come ON. In less than 5 minutes, with the magical power of simple human language instruction, two hardcore activations of “ctrl-C”, and one re=prompt, I have a magical optical recognition AI that feeds me real time analysis of a game I’m playing. The only reason I even saw the code is because I left the FullMonty box checked, otherwise, I spent the evening cutting and pasting each hand and getting a quick ranking.

Why It’s Awesome

Most tech demos are reachy, “see what you can do with this!” best-case scenarios without real world utility, using a narrow, perfect model consumer to get results.

This is the opposite. I’m being a TERRIBLE consumer, and yet the utility is massive. My two prompt Magic Draft Coach sped up my initial “scan and sort” process significantly. After an hour or so of coaching, I absolutely know the set better than if I had spent the same hour or two of “learning games” unaided, looking back and forth between resources, double checking card texts, checking a rule here or there.

All told, the magic box saved me hours. Not the “fun” hours, the actual playing and decision making. No it saved me the boring bits of the hours -looking things up, reading analysis, making connections, so I could have more of the fun bits.

$20 a month for more time doing something I already love, and less time doing something that felt like a chore?

Very few things are that good a trade. Where can you save an hour of your life using ChatGPT? Got a mailing list to organize? Wedding tables to stratify? A project to break down into pieces? How much of your life might you get back, for $20?

Why It Remains Weird And Maybe Not Awesome

Is it a little weird that I “grabbed” thegathering.cc’s rating system (mad props), and screen capped the game? Probably? I’m going to guess I signed something in the blood of a time travelling noble that I wouldn’t do either of those things when I opened up the Yogurt seal this morning, I can’t keep up. By hitting “Crtl-V” I probably asserted unilateral commercial ownership of Hasbro’s IP to OpenAI and now I’m banned from Pawtucket, RI.

And the spontaneous analysis actually caught me off guard. If I asked it about an ancient Magic set, I’d expect it to have it all in its Magic Math People Predictor engine, but this is a new set, likely not spoiled before the 4o training date in 2023. So that was a nice set of intuitive leaps it made to add value to the interaction, unprompted.

But also, that unprompted intuition is probably based on stuff I was writing for websites and magazines in 2003. Which I gotta say, still feels weird,

But, here we are.

P.S. - I acknowledge this a silly and wasteful use of the most foundational technology shift of my already bonkers lifespan, and that I likely consumed the water resources of a small village somewhere in doing so. For which I feel guilt. But hey, at least it’s on brand.