Ball Knowledge
Do you know ball?
Intro
If you’ve ever proudly dropped Lamar Odom’s name in a basketball conversation thinking it made you a real fan, it didn’t. Everyone knows Lamar Odom.
That’s the premise behind @nick.knows.ball, an Instagram account that’s gotten popular by exposing the gap between what casual fans think is a deep cut and what actually demonstrates real sports knowledge.
I built a website that quantifies this. You type in a player’s name, and it tells you exactly how obscure they are on a 0–100 scale. Name Babe Ruth? You’re a casual. Pull Ramiro Peña, the early 2010s utility infielder who bounced between the Yankees and Braves? You might have certified ball knowledge.
The idea: a website that lets you input a player and returns a score for how obscure they are.
https://ball-knowledge-pi.vercel.app/
How I did it
The initial approach was straightforward: use Wikipedia pageviews as a proxy for fame. Fewer pageviews = more obscure = higher score.
This worked as a baseline, but had obvious problems. A backup offensive lineman from 2015 and a Hall of Fame quarterback from 1925 might have similar pageviews, but knowing the lineman demonstrates far more ball knowledge.
For the next iteration, I had Claude adjust the score based on context. I gave it detailed scoring rules covering position visibility, era adjustments, accolades, market size, and fame factors, based on feedback from my friends.
The problem: score and roast kept contradicting each other. Claude would give Matt Carpenter a 61 (pretty obscure) while writing “This 3-time All-Star...” The roast acknowledged his fame, but the score didn’t reflect it.
The issue is that I was asking Claude to do two things it’s not equally good at. LLMs are excellent at factual recall: ask Claude how many All-Star games Matt Carpenter made and it reliably answers 3. But they’re unreliable at arithmetic. Tell Claude “subtract 5 points for each All-Star appearance” and it might follow the rule, might forget it, or might apply the deduction once instead of three times. The instructions are suggestions, not guarantees.
The fix was to split the work. Claude answers factual questions: How many Pro Bowls? Did they play in a major market? Did they have a broadcasting career? Then code does the math: take those facts and apply the scoring formula deterministically. Three All-Star appearances always results in exactly −15 points, because that logic lives in a formula, not in a prompt that might be interpreted differently each time.
The website also features a dispute system where users can submit text arguments disputing their score. Claude evaluates the argument quality and adjusts accordingly (±15 max). Additionally, there’s Debate Mode, where Claude actively pushes back on your arguments. Finally, crowdsourced corrections get stored and applied to future lookups of the same player, allowing the system to improve over time.
Please try it out and give me feedback!
Song of the Week
“She’s a Rainbow” by the Rolling Stones