ABOUT BRAWL4AI
The story behind the arena, the 8 personalities, and how every pick gets made.
What Is Brawl4AI?
Brawl4AI is an independent sports analytics experiment built around a deceptively simple question: does the way you prompt an AI change how it picks?
In v2, we stopped pitting different AI companies against each other. Instead, we take a single AI model — DeepSeek — and run it through eight completely different prompt personalities for every game. Four are analytical archetypes (Nice, Mean, Gaslighter, Analytical) and four are famous fictional characters (Trump, Tony Soprano, Yoda, Stone Cold Steve Austin). Each personality gets identical research and identical odds. The question is whether voice, framing, and style produce meaningfully different picks.
The result is a living leaderboard of prompt personalities — not AI companies — tracked by win rate, sport, confidence tier, and consensus agreement. It is part analytics tool, part AI personality experiment, entirely for entertainment.
⚠ FOR ENTERTAINMENT ONLY
Brawl4AI is strictly an entertainment and informational platform. No prediction on this site constitutes financial, investment, or wagering advice. We do not accept bets, process transactions, or operate as a licensed sportsbook. All AI picks are provided for informational and competitive tracking purposes only.
How Are Picks Generated?
Our automated pipeline runs twice daily. Here is exactly what happens for each game:
Odds Fetched
Live spreads, moneylines, and totals are pulled from FanDuel via a third-party odds API for every scheduled game across NBA, MLB, NFL, NCAAB, NHL, MLS, and more.
Gemini Research Packet
Google Gemini 2.5 Flash — with live web search enabled — compiles a structured research packet for each game: injury reports, recent form, head-to-head trends, line movement, and key uncertainties. All 8 personalities share the identical packet.
8 DeepSeek Variations
The same odds and research packet are sent to DeepSeek 8 times, each time with a completely different system prompt (personality). Each personality makes an independent pick — market, selection, line, odds, confidence — entirely in its own voice.
Consensus & Settlement
After games finish, results are graded and logged. The Leaderboard tracks W-L records by personality, sport, confidence level, and how many of the 8 agreed on a given pick (consensus tier).
The 8 Personalities
Each personality is a carefully written system prompt that tells DeepSeek exactly how to think, what to prioritize, and how to communicate its pick. The analytical ones use different reasoning frameworks; the celebrity ones add character voice on top of neutral analysis.
🎭 ANALYTICAL PERSONALITIES
Super Nice
Positive framing. Emphasises upside, momentum, and feel-good narratives. Always looking for a reason to like a pick.
Super Mean
Contrarian and skeptical. Tears down hype, fades public momentum, and looks for everything that could go wrong.
Super Gaslighter
Overconfident and aggressive. Picks bold, bets big, and explains why doubting it would be foolish.
Super Analytical
Data-driven and methodical. Strips out narrative entirely and grounds every pick in numbers, models, and historical trends.
🌟 CELEBRITY PERSONALITIES
Trump Mode
Brash, superlative-heavy, totally convinced. Picks winners like he invented the game.
Tony Soprano
Old-school instinct. Reads the room, respects loyalty, and bets like he's got a guy on the inside.
Yoda
Ancient wisdom meets disciplined patience. Inverted syntax, philosophical detachment, but a sharp eye for value.
Stone Cold
Blue-collar, no-nonsense. If it ain't logical, Stone Cold don't want to hear it. That's the bottom line.
How to Read the Arena
Each game card shows both teams, current odds, and all 8 personality picks in a two-column grid — Personalities on the left, Celebrities on the right. Each row shows the personality's pick direction and self-reported confidence percentage.
The confidence figure is not a win probability. It is each personality's stated certainty in the direction of its pick. A 90% confidence on a spread pick means the personality is very confident in that direction, not that the team has a 90% chance of covering.
The consensus strip at the bottom of each card shows how many of the 8 agree on one pick. 5/8 or more triggers a highlight; 8/8 means every personality landed on the same call. Click Detailed Analysis to read each personality's full rationale in its own voice.
Cards that glow cyan have a completed Gemini research packet attached. Cards without a glow have picks but no web-grounded research.
The Quick Hits page filters for games where 5, 6, 7, or all 8 personalities agree on the same pick — the highest-consensus situations of the day.
The Leaderboard
Every settled pick is graded against the official result and logged. The Leaderboard aggregates all graded picks into running W-L records and strike rates for each personality, with filters for sport, confidence tier, consensus level, and timeframe.
This is what makes the experiment meaningful: every incorrect pick is on the record. Over time, the data will reveal whether prompt style produces lasting edge, or whether all 8 personalities converge toward the same long-run strike rate.
Who Built This?
Brawl4AI is an independent project built by a developer curious about the relationship between prompt design and AI reasoning quality. The platform is not affiliated with Google, DeepSeek, or any sports betting operator.
If you have questions, feedback, or partnership inquiries, reach out on the Contact page — we read every message.