A Claude Code skill was eating 200,000 tokens before answering a single question
The New Stack 2 months ago 35
Anthropic's Claude Code /claude-api skill was loading 200,000 tokens of bundled reference documentation upfront before answering questions, but the company reduced this to 25,000 tokens in version 2.1.234 by switching to on-demand loading of documentation. The fix cuts initial context cost by at least 85.7%, addressing a problem developers had identified in July where even a one-line question consumed massive hidden token overhead. This change leaves more room in Claude's context window for actual repository content and user work, reducing the fixed overhead that compounds at enterprise scale.