Using Claude to Extract Data from a 1997 Football Manager Game
Ben Nuttall reverse-engineered FIFA Soccer Manager 97 (FSM97) using Claude as a data extraction assistant. Pointing Claude at the Wine-installed game directory, the AI located SM97.DAT and parsed player names, stats, club assignments, stadiums, and manager data — all mapped to real-world football entities. The goal was reproducibility: the entire pipeline now lives as Python scripts on GitHub.
Key Details
- Source file:
SM97.DATcontains all game data including player full names (e.g., "David Beckham" vs. in-game "D. Beckham"), stadium names, club nicknames, and even managers never shown in game. - Initial query: Claude answered simple questions (biggest stadium, highest-rated player) by reading the binary file directly.
- CSV export: Claude generated CSV files for all data; column names for player stats were unknown, so Nuttall launched the game to map them (e.g., using David Seaman's stats as a calibration reference).
- Site generation: After fixing a few hallucinated data points, Claude built an interlinked HTML site at fsm.bennuttall.com showing players, clubs, stadiums, and trivia.
- Data corrections: Abbreviated team names like "Sheffield W" were expanded to "Sheffield Wednesday". Stadium typos (e.g., "Bramall Lane Ground") were fixed without altering gameplay data.
- Easter eggs discovered: Olympic champion Daley Thompson appears as a player at Mansfield Town. Player-manager relationships detected (e.g., player also listed as manager for same club).
- Shared stadiums: Clubs like Crystal Palace and Wimbledon both use Selhurst Park; the data now tracks this.
- Top stats pages: Generated lists of top-rated players, best players by age group, top player-managers, and stadiums by capacity. Notably, the "EA All Stars" club contains fictional highly-rated players.
Reproducibility
All Python code to extract data and build the website is open source on GitHub. Others can run the same pipeline without needing Claude or any AI tool by using the published CSVs and scripts.
📖 Read the full source: HN AI Agents
👀 See Also
Comparison of 8 AI Coding Models on Real-World TypeScript Feature Implementation
A developer tested 8 AI coding models on implementing a /rename command in an open-source TypeScript Telegram bot project, evaluating them on cost, execution time, correctness, and technical quality. GPT-5.4 scored highest on implementation correctness while GLM 5 offered the best cost-performance ratio.
Claude Desktop Feature Request: Session Start Hook for Automatic Initialization
A developer building persistent context systems for Claude Desktop identifies a gap: the User Preferences field only injects instructions when the user sends the first message, requiring manual triggers for initialization. They propose adding an "On Session Start" execution field that runs automatically when a new conversation opens.
Introducing cltree: A File Tree TUI for Claude Code
cltree is a split-pane TUI that displays your project file tree in real-time alongside Claude Code, showing the current working directory, hiding noise, and allowing all keystrokes to pass through.
Tastebud Memory: Reversible Agent Memory via Hyperdimensional Computing Vectors
Hyperdimensional computing replaces vector search for complete recall: list ALL days touching a project, detect unnamed workstreams, and decompose daily logs losslessly via dot products.