Claude Opus 4.6 system prompt consumes 60-70% of context

This is less a bug than a design tax — feature-rich AI assistants pay for their capabilities in context overhead, and most users never see the bill.

• –System prompt bloat is a known tradeoff: 20+ tool definitions, safety policies, artifact rules, and skill descriptions stack up fast before a user types a single word
• –The 30-40% headroom is comfortable for casual chat but tight for agentic workflows, deep research sessions, or large file analysis
• –Developers building on the Anthropic API with minimal system prompts get substantially more usable context than consumer Claude.ai users
• –This also explains why context window size comparisons between models can be misleading — raw token count doesn't account for how much the vendor pre-loads