Universal CLAUDE.md trims Claude output 63%

Clever idea, but the real value here is less “63% fewer tokens” than “less annoying agent behavior at scale.” It is a practical prompt-shaping tool for high-volume Claude workflows, not a universal cost reducer.

• –The benchmark is explicitly directional, not controlled, so the 63% figure should be treated as a signal, not a guarantee.
• –The repo is unusually honest about the trade-off: `CLAUDE.md` adds input tokens on every message, so short sessions can cost more, not less.
• –The strongest use case is repetitive automation: code review, agent loops, and other output-heavy tasks where verbosity compounds.
• –It targets the stuff developers actually complain about: sycophantic openers, hollow closings, restated prompts, Unicode noise, and scope creep.
• –If you need strict parseability or hard behavioral guarantees, structured outputs and tool schemas are still a better foundation than prompt rules.