Skills
Claude Code Instruction Slot Budget: Frontier Models Reliably Follow ~150 Discrete Rules, Claude Uses ~50
LLMs process instructions as positional slots, not memory—frontier models reliably follow 150-200 discrete instructions per context, and Claude Code's own system prompt consumes ~50 of those slots leaving 100-150 for user rules. Exceeding this budget doesn't produce errors; it silently drops compliance on later rules. The fix is fewer, higher-priority, code-enforced rules rather than exhaustive prose checklists.
Source
↳ Follow the thread