I Built The Token Saver Skill To Cut My Token Use By 90%. Here Is What It Can And Cannot Do For You.

Source: Nate’s Newsletter Date: 2026-07-29 URL: https://natesnewsletter.substack.com/p/reduce-ai-token-usage

Summary

The author measured that 95.73% of their AI input tokens across 3.77 billion tokens were reused context — prior exchanges, files, screenshots, rejected answers carried forward by default — rather than newly typed material, and built a “Token Saver” skill enforcing four concrete changes in Codex and Claude Code to trim that carryover. The piece is explicit that reused context is often necessary for continuity, not waste to eliminate outright, and that validated final savings numbers require a comparison the author hadn’t yet published at time of writing.

Implications

• Context/token efficiency: names the actual driver of cost growth in long agent sessions — accumulated re-sent context, not verbose prompts — which bears directly on how long-running or resumed agent sessions should be structured to bound cost.
• Agentic engineering patterns: packaging cost discipline as an installable “skill” enforced at the tool level (rather than relying on habit alone) is a distribution pattern worth tracking as skill ecosystems around Codex/Claude Code mature.