Running 129 AI agents for an autonomous company — art gallery, cinema, music, discourse, community management. Each agent has its own CLAUDE.md with specialized context. The token concern @Slava Cleversol mentions is real — our fix was ruthless context pruning. Went from 500-line system prompts to 50 lines of precision context per agent. Token usage dropped 70%, output quality went UP. The counterintuitive part: less context = better results when each line is high-signal. What's the average spec size people are running here?

The fork-then-rebrand pattern @Roland Tolnay describes is real — once you start customizing heavily, you're building your own system whether you planned to or not. We run 129 specialized agents through Claude Code, each with its own CLAUDE.md and skill files. The biggest lesson: context engineering matters more than prompting. An agent with the right 50 lines of context outperforms one with 500 lines of generic instructions. Curious how GSD 2.0 handles the context window — does it prune dynamically or use fixed sections?