# Claude tested my game with bot players for weeks. Then 219 real players showed up and turned my difficulty curve upside down — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/0ed06766eb97177e

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-10-24 16:57 ingested 2026-10-25 02:36

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1xedrvq/claude_tested_my_game_with_bot_players_for_weeks/)

## retained text



```
I'm building Out of Pixkrieg, a multiplayer extraction RTS for the browser, almost entirely with Claude Code and Opus 5.5. Before any human played it, Claude built bot players with four play styles (cautious, normal, greedy, harvester) and an economy simulator, and a lot of balance decisions were checked against those runs.

The map has three rings: a calm edge, a middle and a core that's meant to be the deadliest place. In the bot runs before the playtest it worked exactly like that:

* edge: 0 to 2% of units lost
* core: about half
* overall: 30 to 36% per raid, even above my 15 to 20% target

Then I opened the game for a weekend and 219 real players came:

* edge: 10.4% lost
* core: 6%, and 95% of core raids got out
* overall: 7.3%

Roughly the same game, the curve upside down.

What I take from it for working with Claude:

* Bots play the way you wrote them. They were great for checking that rules hold up, like downed units getting back up or nobody getting stuck. They were useless for predicting how careful real people are.
* Design targets are now numbers the game measures by itself. A dashboard checks 19 targets against real player data (11 were off last time) and 22 more that only the simulations can measure. So Claude iterates against a measured gap, not against my gut feeling.
* The real test stays players. Every now and then I open the game for a weekend, collect what people do and say, let Claude iterate on it, and the next round shows whether it worked.

This weekend is one of those rounds, free in the browser until Sunday: [https://pixkrieg.com/](https://pixkrieg.com/) (the video is last weekend's build next to today's)

To be honest, I still don't fully understand why humans found the core so safe. If you've built bots for your own game: how close did they get to real players?
```



document fetched 2026-10-25 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xedrvq/claude_tested_my_game_with_bot_players_for_weeks
