2026-06-26

Claude as a Reverse-Engineering Assistant for a 1997 Football Manager Game

What Ben Nuttall did here is oddly compelling: he took Claude and pointed it at the dusty innards of a Windows 95-era football management game, then used it to turn an opaque binary blob into something navigable, queryable, and even publishable on the web. For Claude and Claude Code users, this is the sort of project that feels genuinely useful rather than demo-cute — a mix of archaeology, data wrangling, and careful human verification.

Key Points

Ben revisited FIFA Soccer Manager 97, a game he played as a child, and found it still runs in Wine on Linux.
He asked Claude to inspect the game files and focus on real-world football data such as players, teams, and stadiums.
Claude found the relevant data in SM97.DAT and could answer simple questions quickly, like the biggest stadium or best-rated player.
He had Claude generate an HTML summary page, then CSV files so he could inspect and verify the extracted data.
When Claude didn’t know some player-stat column names, Ben opened the game and wrote them down manually to calibrate the extraction.
He expanded the extraction beyond English leagues to include all players and clubs.
To avoid hallucinations or re-parsing the binary every time, future queries were run against the extracted CSV files.
He then asked Claude for a more comprehensive interlinked website, and also published the extraction code on GitHub so others can reproduce it.
He corrected some data, including full club names like “Sheffield Wednesday” instead of shortened in-game forms, plus some stadium names and other odd entries.
The game contains data that is not actually used in gameplay: full player names, manager names, nicknames, stadium names, and sometimes town/address data.
Some of the data reveals fun oddities, like Peter Shilton at Leyton Orient and Daley Thompson at Mansfield Town as an easter egg.
The extracted data also captures shared stadiums and player-managers.
He built special pages for things like top-rated players, best players by age group, top player-managers, and stadiums by capacity.
A notable curiosity is the EA All Stars club, which includes developers and staff rated among the game’s best players.
Ben explains one player-record example in detail, showing how a player’s 87-byte record maps to name fields, skills, and date-of-birth data.
He describes the so-called 2079 bug: player birthdates are stored as a 16-bit little-endian unsigned integer counting days since 30 December 1899, so the game overflows after 5 October 2079.
The game’s FMV cutscenes use EA’s TGQ format, and ffmpeg can decode them with the ea demuxer.
He was able to extract and document the video files, though he says he wasted time trying to extract stadium graphics without success.

My Take

What strikes me is that this is exactly the kind of Claude project I’d actually want to do: not “write me a blog post,” but “help me make sense of a weird binary format I don’t fully understand yet.” That’s where an LLM feels genuinely strong. It can pattern-match the structure, propose a parsing model, and accelerate the first 80 percent of the work. But the important part here is that Ben didn’t stop at “Claude said so.” He verified fields by hand, corrected names, and moved the final workflow onto CSVs so the system became reproducible instead of vibes-based.

I think that reproducibility point is the biggest win. A lot of AI-assisted reverse engineering stories quietly collapse into “the model guessed correctly once.” Here, the interesting move was to turn Claude’s exploratory work into an ordinary data pipeline. That’s the boring part, and also the part I trust. If I were doing this with Claude Code, I’d follow the same shape: let the model inspect, summarize, and draft parsers, then freeze the output and interrogate that, not the source binary, for the rest of the project.

I also like that the article doesn’t pretend the LLM did everything. Ben still had to launch the game, note down stats, spot weird naming quirks, and make judgment calls about what to correct. That feels honest. The model is a very good assistant here, not a magical archaeologist.

The EA All Stars bit is the most delightfully human part of the whole thing. Someone at EA clearly had fun with the data. And honestly, that’s the sort of easter egg Claude is good at surfacing because it can connect a weird club name to a hidden pattern across the dataset. I’d be curious whether more old games have this kind of embedded metadata waiting to be unpacked.

The only part that feels a little overhyped, if I’m being blunt, is the idea that Claude “extracts data” by itself. It can help a lot, but the quality comes from the surrounding process: validation, calibration, and deciding what counts as canonical. Claude is the accelerator. Ben did the engineering.

The takeaway is simple: Claude can be surprisingly effective as a reverse-engineering copilot, especially when the end goal is a clean, shareable dataset. But the real value comes when you use it to build a toolchain you can trust, not a one-off answer.