Getting Claude to extract data from a 1997 football manager game

Recorded: May 27, 2026, 11:03 p.m.

Original

Summarized

Ben Nuttall - Getting Claude to extract data from a 1997 football manager game

Ben Nuttall

Home
Blog
Talks
Projects
About

Blog » 2026 » April
Published 29 April 2026

Getting Claude to extract data from a 1997 football manager game

There's a football manager game I got for my birthday when I was nine years old: FIFA Soccer
Manager
97. I spent countless hours playing it as a child, trying to win the league, the cup or just avoid
getting the sack! Two years later I got a newer manager game with better graphics and more complex
gameplay but it ran incredibly slow and I didn't enjoy playing it at all, so I stuck with the 1997
one.
Every couple of years I get it out again and spend an evening basking in nostalgia and reliving the
experience. Originally it was made for Windows 95, and it was still working on Windows until (I
think) Windows 10. But it works perfectly well in Wine on Linux, so I can still play it today. It
just has a teeny tiny resolution and looks silly on my 42" monitor.

I wondered if I could get Claude to extract data from the game files. I pointed
it at the directory my Wine installation installed to. I told it I was only interested in players,
teams, stadiums and that kind of thing - the stuff that maps to the real world of football rather
than internal game data. It found everything in the file SM97.DAT and was very quickly able to
answer simple questions like which is the biggest stadium or who is the best rated player.
I asked it to make an HTML page summarising all the data. It produced this:
https://files.bennuttall.com/fsm97/
I asked it to create CSV files of all the data it found, so I could make sense of it and make sure
it looked right. There were a few issues with it messing up the data but it was easily able to fix
it. It didn't know what all the column names for the player stats were, so I launched the game and
wrote them all down along with David Seaman's stats so it could calibrate. I expanded the scope to
all players and clubs, not just the English leagues.
Once I was happy with the CSV files, all future queries would be scripted against these files so I
could be sure it was coming from extracted data and not extracting from the game data again or even
hallucinating. I then asked for a more comprehensive website of all the data, with lots of
interlinking. I was really impressed with what it put together, and I spent some time diving deeper
into the data and making tweaks to the website.
I wanted to produce a set of Python code to make this process reproducible, so if anyone else wanted
to do it they could do so without needing an AI tool of their own. It's now on
GitHub, and I've published the website too:
fsm.bennuttall.com

fsm.bennuttall.com

I've worked with Claude to fine-tune the way the website works, trying to interlink everything and
make it easy to find what you're looking for, or explore the data looking for interesting things.
Stats, trivia and EA All Stars
I always found it annoying that it only ever used short versions of team names like "Sheffield W"
instead of "Sheffield Wednesday",
purely to keep the strings short enough to use everywhere. I had Claude correct them to their proper
titles. Stadium names are not used in the game, but they all exist in the game data, albeit with odd
mistakes sometimes, like "Bramall Lane
Ground". I corrected those too. Some more
obscure club or stadium names were just wrong so I fixed those too. I didn't want to mess with too
much of the game data, but felt these changes were reasonable.
One thing I found interesting was that there's data in there that isn't used in the game at all.
Players go by first initial and Surname, like "D. Beckham", but the data has "David
Beckham". It has all the
manager names of all English leagues and a few others, but they're never used in game. Same for club
nicknames and stadium names (and sometimes town/city and even first line of address!).
This revealed oddities like P. Shilton in the game actually being former England goalkeeper Peter
Shilton, who really did play for
Leyton Orient until 1997, at the age of 47;
and oddly, Olympic decathlon champion Daley
Thompson at Mansfield
Town (an easter egg).
I didn't realise there were some clubs that shared stadiums, which is reflected in the game. I
wonder if you managed Wimbledon and expanded the
stadium (Selhurst Park), if you played against
Crystal Palace away, would you see the ground at
its default appearance or would you in fact see your own upgraded home stadium? (The Reddit
community confirms they do appear as different stadiums!)
I made sure the extracted data was aware of the concept of shared stadiums, and it also picked up on
the fact that some of the clubs' managers were also listed as players for the same club - a once
popular "player-manager" role.
I put together a few special pages showing some interesting stats, such as top rated
players, top 10 best players in each age
group, top rated
player-managers and stadiums by
capacity.

One thing that stands out in top player stats is a set of highly rated players from a club called
"EA All Stars". None of these are real players —
they're actually the game developers and other staff who worked on the game. The Assistant Producer
Mark Bergan made himself one of the
best rated players in the whole game, rated as highly as David
Seaman, George
Weah and
Romario.

EA All Stars

Data extraction method
The data extraction method
is explained in detail on GitHub. Here's an example:
David Seaman is the first player record in Arsenal's block. His raw 87 bytes:
0: 44 61 76 69 64 00 00 00 00 00 00 00 00 00 00 00 David...........
16: 00 00 00 00 00 00 00 00 53 65 61 6d 61 6e 00 00 ........Seaman..
32: 00 00 00 00 00 00 00 00 00 00 00 1a 00 00 01 48 ...............H
48: 5e 46 47 4a 48 10 48 19 2f 1a 57 5d 2c 2a 18 57 ^FGJH.H./.W],*.W
64: 57 5f 1d 50 58 40 23 00 00 50 b6 52 01 e9 5a 04 W_.PX@#..P.R..Z.
80: 04 01 ff 8a 05 00 00 .......

Names

Bytes
Hex
Value

[0:6]
44 61 76 69 64 00
David

[24:30]
53 65 61 6d 61 6e 00
Seaman

Stats block (starting at byte 42)
Byte 42+1 = 0x1a = 26 → nationality index 26 (England)
Byte 42+2 = 0x00 = 0 → position GK
Byte 42+4 = 0x01 = 1 → shirt number 1

Skills (stats[5:28] = bytes 47–69)
48 5e 46 47 4a 48 10 48 19 2f 1a 57 5d 2c 2a 18 57 57 5f 1d 50 58 40

Attribute
Hex
Value

Speed
48
72

Agility
5e
94

Acceleration
46
70

Stamina
47
71

Strength
4a
74

Fitness
48
72

Shooting
10
16

Passing
48
72

Heading
19
25

Control
2f
47

Dribbling
1a
26

Coolness
57
87

Awareness
5d
93

Tackling Det.
2c
44

Tackling Skill
2a
42

Flair
18
24

GK Kick
57
87

GK Throw
57
87

GK Handling
5f
95

Throw-in
1d
29

Leadership
50
80

Consistency
58
88

Determination
40
64

Remaining stats
stats[28] = 0x23 = 35 → Greed
stats[30] = 0x00 = 0 → Form
stats[31] = 0x50 = 80 → Energy
stats[35] = 0xe9 = 233 ⎤
stats[36] = 0x5a = 90 ⎦ → 233 + 90×256 = 23273 days → 1963-09-19 (DOB)

Age on 29 July 1996: 32.
The 2079 Bug
I'd read on the thriving Reddit community for this game about the
"2079 bug". Apparently, if you play the game through to 2079, at some point the age of players hits
an overflow and they all think it's time to retire.
This is because each player's date of birth is stored as a 16-bit little-endian unsigned integer:
the number of days elapsed since 30 December 1899 (the same base date used by Microsoft Excel and
the game's internal date system). The two bytes are combined as:
days = stats[35] + stats[36] * 256
dob = date(1899, 12, 30) + timedelta(days=days)

This is a WORD (16-bit), so the maximum representable date is day 65,535 — 5 October 2079. After
that point the game overflows and all player ages are calculated incorrectly. This is known as the
2079 Bug.
I've never played the game that far. I have occasionally taken it to the current day (as recently as
the 2020s), but it's known that the game has limitations in endurance.
Full motion video cutscenes
The game features a number of full motion video cutscenes. These include specially-filmed
live-action football footage, heavily tinted blue/purple, with overlaid title text in a
typewriter-style font. They were shown when you won or lost a cup final, won the league, got
promoted, relegated, or sacked. There was also an intro and a closing credits video.
I knew these videos would be in the data too, though I wouldn't have known what format or how they'd
be stored. Claude managed to find and identify them:

The game's video files use EA's proprietary TGQ format (.TGQ extension), also known as EA TGQ or
EABT. The container uses EA's chunk format, identifiable by the SCHl magic bytes at the start of
each file.

I hoped it would be able to maybe extract the frames, but it had no problem decoding the video
files.
ffmpeg can decode these files natively using the ea demuxer:
ffmpeg -i INPUT.TGQ -c:v libx264 -c:a aac OUTPUT.mp4

Docs and explanation of the video files can be found here:
github.com/bennuttall/fsm-97-data/blob/main/docs/videos.md

Credits
The end credits are something of a masterpiece. The game developers and all the staff who worked on
the game are credited, along with childhood photos and aspects of what they worked on, including
example lines of C/C++ code, stadium sketches, colours and coordinates, specs and more:

Since many of these staffers are also included in the game, I was able to create a credits
page listing them and link to their player listings on the EA
All Stars.
Join the project
If you have a copy of the game, head over to the GitHub
page and see if you can extract the data yourself, and have a play with it, maybe find things I've
not found yet.
There are a few other things I'd like to investigate. I wasted a bunch of time (and Claude usage)
trying to extract the stadium graphics, to no avail. I was hoping to extract one of these for each
club:

Highbury

The end?

Tags: python
claude
ai
web development
github
open source
football

« Glen Etive

Archive

All
Tags

2026

2025

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

2006

Blog

Blog index
Archive
Tags
Images
Atom feed

Find me online

Mastodon
GitHub
LinkedIn
BlueSky
Twitter
Flickr
YouTube

This site

Ben Nuttall undertook a project focused on extracting meaningful real-world football data from the game files of FIFA Soccer Manager 97, utilizing the capabilities of the large language model Claude. The initial motivation stemmed from a personal interest in the game and the desire to systematically analyze its contents, focusing specifically on elements mapping to actual football entities such as players, teams, and stadiums, rather than internal game mechanics. Nuttall directed the AI to analyze directory files, including SM97.DAT, successfully identifying this external data. The AI initially proved effective at answering simple queries regarding the game, such as identifying the largest stadium or the best-rated player, and generating summary outputs in formats like HTML and CSV files.

When processing the extracted data, Nuttall encountered necessary corrections, as the AI sometimes produced erroneous data. This required manual intervention to calibrate fields, such as determining the specific column names for player statistics, by referencing the game itself. The process was then refined by scripting subsequent queries against the generated CSV files to ensure that all derived information originated strictly from the extracted data, thereby preventing hallucination or re-extraction from game files. Nuttall further expanded the scope to include all players and clubs, not limited to the English leagues, and developed an interactive website with extensive interlinking, leveraging Claude to fine-tune the structure and navigation for enhanced user exploration of the complex dataset.

During the data refinement, Nuttall discovered subtle inconsistencies embedded within the game data. For instance, the game often used abbreviated team names, and stadium names contained discrepancies, such as "Bramall Lane Ground." Nuttall corrected these textual errors, and further analysis revealed unused data points, including manager names and club nicknames, which offered historical context. This led to the discovery of unusual in-game data points, such as the appearance of historical figures like Peter Shilton and Daley Thompson in contextually odd situations, and the reflection of shared stadium concepts within the game’s structure. Furthermore, an investigation uncovered statistical anomalies, including a specialized set of highly rated players labeled "EA All Stars," which included the game developers and staff members, suggesting embedded meta-data within the game itself.

A deep technical analysis of the data extraction method revealed the underlying structure of the file. Nuttall demonstrated byte-level extraction of player names and statistical attributes, illustrating how names and numbers are encoded in the raw file format. The process involved calculating player dates of birth by interpreting 16-bit fields representing the number of days elapsed since a base date to determine the player's age. This analysis highlighted a critical limitation in the game's design, known as the "2079 Bug," where player ages overflow after reaching a certain limit, demonstrating inherent endurance limitations within the software.

Moreover, the project extended to media extraction, where Claude assisted in identifying proprietary full motion video cutscenes stored in the game using the EA TGQ format. While the AI did not extract the visual frames directly, it helped pinpoint the file structure, allowing subsequent use of tools like ffmpeg to decode the video content natively from these files. The project culminated in the creation of reproducible Python code and a public website, demonstrating a methodology for systematically analyzing legacy game data through the combined application of data mining techniques, programming, and artificial intelligence.