Skills
Cursor cut agent token cost 7% by trimming its system prompt 66% and loading rarely used tools on demand
Cursor's 23 Sep engineering post reports a 66% smaller system prompt with no quality loss after it deleted DO NOT rules that newer models no longer need, and it validated the cut with A/B tests on production traffic rather than evals. Keeping only the read, search, edit and shell tools static and loading the rest on demand cut static tool-description tokens by 60%. Separating stable layers from the growing conversation with explicit cache breakpoints cut cold cache misses by 20%, and numbering every tenth line in file reads instead of every line saved 1.6% of cache-read tokens. All four changes carry over to a home-built harness, and the line-numbering change is the cheapest one to try first.
Source
↳ Follow the thread