Developer trims AI tool descriptions by 40%, trigger rate holds steady in A/B test

A developer running 114 AI skills in Claude Code discovered that exceeding the system prompt's character budget silently strips tool descriptions, leaving only names that the model cannot act on autonomously. The platform allocates just 1% of the context window — roughly 8,000 characters on a 200K-token session — for skill listings, and overflows are evicted starting with least-used tools, creating a compounding invisibility loop. After decompiling the local binary to confirm the actual budget rules, the developer rewrote 41 tool descriptions to under 250 characters each, reducing total description text from 44,775 to 34,375 characters across 116 skills. Behavioral tests showed no meaningful drop in trigger rate — 90 out of 96 passed versus 88 previously — confirming that leaner descriptions preserved tool discoverability. The findings suggest that developers with large AI toolboxes should audit description lengths and consider overriding the default budget variable before drawing conclusions from zero-usage metrics.