Brass & Iron. A football GM game that plays like building a team, not filling in a spreadsheet.

I design every screen and rule, and decide what ships. Claude writes the code to my specs and runs the tests I set. It's at version 24 now, and every balance change is measured.

My roleDesign, product, direction
TeamMe, with Claude writing the code and running my tests
Versionv24.8, in progress
6–15 → 1–5actions to find the team's number, before and after the fixes from the blind test
1 → 4 of 5testers who could find out why a game was lost
13.0 vs 8.1wins a season for a good GM against an idle one, on the same leagues
01

The problem: every GM game is a spreadsheet

Football GM games are deep, and most of them look like a database. I wanted the same depth in a game where you see people, a place and a season, and can still find every number.

Before · v13.7Coaches as abstract shapes and rows of numbers
After · v24.8Coaches as people, each with a face, a trait and a fit with your plan
02

How I work: I design and decide, Claude builds to spec

Every feature runs the same loop: rules in text, a sketch I judge, a final, the build, then a measured balance check. Seven written documents carry the decisions between chats, because the AI forgets and the docs don't.

  1. 01BriefWhere we are, what to read
  2. 02RulesDecided in text first
  3. 03SketchGrey boxes, no colour
  4. 04FinalFull visual, dark and light
  5. 05BuildOne version per feature
  6. 06MeasureTwo bots, same leagues
  7. 07Hand-overUpdate the brief, commit

Rules before screens came from a mistake. The Staff screen took seven sketch rounds because its rules were decided while sketching. Once rules came first, each board took one round.

BeforeAfter
03

The work: rows into people

Four places where a table became something you'd want to open.

Hiring as portrait cards. Coaches are people you choose between, each with a scheme, a fit and an asking price. v24.6
Game day, from the press box. You watch your game play out, with the stats and leaders beside it. v24.8
The first day. The owner sets your goal in his box, then Gus starts handing you tips. v24
Choosing a club. Each of the 32 clubs has a city, an owner, a rival and a history to live up to. v24
The Desk. The GM's office. The owner calls, and one move a week.
Scouting. 300 prospects, and what your scouts actually know about each.
The roster. The table is still there, for when you want every number.
The start screen. The name, drawn in pixels over a city at night.
04

Testing: blind testers, real logs

Five simulated players played the game without ever seeing the docs: three who don't follow football and two who play management games. Every turn was logged with a screenshot, a thought and an action. A separate agent scored the logs blind. I fixed what they got stuck on, then ran the same tasks again.

THE SAME TASKS, TWICE10 sessions each
Taskv22v23, after the fixes
Find the team's number6–15 actions1–5 actions
Find out why we lost1 of 5 testers4 of 5 testers
Before · v22The score and every drive, but the space over the scoring plays (ringed) says nothing. One tester: "Scoring list is cut off and won't scroll."
After · v23One line over the scoring plays: "Lost it on 1 giveaway to 0", then what else decided it.
05

How I measured this

Each number above comes with how it was made, and what it can't tell you.

The two bots

Wins must be earned

A good GM and an idle GM play the same 8 leagues for 10 years each. The target is the gap between them. On v24 the good GM wins 13.0 games a season, the idle one 8.1, and the idle one is fired in 7 of 8 careers.

The blind testers

A walkthrough with a log, not user research

The testers are simulated, not people. They're good at showing what's hard to find and poor at showing how it feels: an agent can only pretend not to know football. Real players come next.

06

Looking back: the hard part moves from making to deciding

The written system kept the quality up, not the prompts. Next time I'd test with real people sooner.

Design, product and direction: Milos Facius Code and test runs, to my specs: Claude Screens from v24.8 unless marked. The web build you can play is v20.