Token Saving Skill

Every turn of an agent session resends the entire conversation — most AI spend is invisible waste: bloated instruction files, full-file reads, verbose output, one flagship model for everything. Token-Saving is a platform-agnostic skill that fixes this at the source. Load one small SKILL.md and your assistant applies eight technique areas automatically: context hygiene, prompt caching, model routing, output discipline, session compaction, subagent delegation, waste-pattern avoidance, and usage measurement. Included and tested: - 3 automation scripts — context-threshold guard, 90% budget alert with cheaper-model nudge, per-call usage logger (CSV) - 5 prompt templates — terse output style, diff-only edits, compressed subagent returns, instruction-file compression, model routing - Cost math reference — a worked example showing the same 40-turn session going from ~$7.65 to under $1 - Platform guides — Claude Code, Cursor, direct API, LangChain The package follows its own advice: the always-loaded core is small; the heavy material loads only on demand. Works with any AI assistant that can read files. No dependencies beyond bash and jq for the optional scripts.