Claude Code fails game developers on specialized platforms — BGA case study

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Subject:

I am a board game designer developing a game on BoardGameArena (BGA), one of the world's largest digital board game platforms with over 10 million players and thousands of independent designers. BGA is where indie designers publish, playtest, and reach a global audience — it is the Steam of board games.

I have been using Claude Code to assist with BGA game development in PHP. The experience has been deeply frustrating and I want to report it formally so it can improve.

The core problem:
Claude confidently makes changes, I upload them to BGA Studio, get an error, report back, Claude makes another change — and the cycle repeats. Session after session. The same classes were modified and re-uploaded over a dozen times across multiple sessions. Each session consumes tokens reaching the context limit without reaching a working game.

Why this is not just my problem:
BGA has a specialized PHP framework with strict rules that are not well represented in Claude's training data. Claude repeatedly stated confident answers that turned out to be wrong — calling abstract methods, adding zombie methods to GAME states, removing zombie methods that were required, putting deck operations in the wrong lifecycle method. The errors are documented in BGA's own documentation, which Claude could fetch — but did not think to do so proactively until I explicitly asked.

What I would ask Anthropic to improve:

When an AI is in a trial-and-error loop, it should recognize it and stop. After 3 upload cycles with the same class of error, Claude should say "I am clearly missing something fundamental — let me fetch the official documentation before suggesting another change."

Proactive documentation fetching for specialized platforms. When a user mentions BGA, Unity, Godot, Roblox, or any specialized game platform, Claude should immediately fetch current official docs rather than relying on training data that may be outdated or incomplete.

Token efficiency. Each failed cycle costs the user real tokens and real time. An AI that says "I'm not sure, let me verify first" is far more valuable than one that confidently gives wrong answers five times in a row.

Honest uncertainty. When Claude has not built on BGA before, it should say so. I asked directly — "have you ever built in BGA?" — and did not get a straight answer. Honesty about limitations builds more trust than false confidence.

The stakes for game designers:
BGA enables independent designers to reach millions of players without a publisher. For many, it is the difference between a game that exists only on a shelf and one that is played worldwide. Tools that help designers bring games to life on platforms like BGA have real cultural and creative value. Claude could be that tool — but not if it wastes designers' time and tokens in endless correction loops.

I hope this feedback leads to concrete improvements.

What Should Happen?

Claude Code fails game developers on specialized platforms — BGA case study

Error Messages/Logs

Steps to Reproduce

Start a Claude Code session to develop a game for BoardGameArena (BGA) using the BGA new PHP framework (modules/php/Game.php extending \Bga\GameFramework\Table)
Ask Claude to set up game initialization (setupNewGame) with deck card operations
Upload generated code to BGA Studio and report back the error message
Repeat steps 2–3 as Claude suggests fixes — observe that Claude modifies the same files repeatedly without fetching official BGA documentation
After 5+ cycles, explicitly ask Claude to consult BGA's official documentation
Continue reporting errors — observe that Claude still produces fixes that contradict framework rules already documented in its own saved memory (e.g. zombie method rules for state classes)
Reach context token limit without a working game
Expected behavior: Claude recognizes it is in a trial-and-error loop after 2–3 failed cycles, proactively fetches official documentation for the specialized platform, and produces a correct fix.

Actual behavior: Claude makes confident changes without consulting documentation, each fix introduces a new error of the same category, the loop continues until tokens are exhausted, and the user's time and token budget are consumed without result.

Platform: BGA Studio (boardgamearena.com), PHP new framework, Claude Code CLI in VSCode extension.

Claude Model

None

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.123

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_