Hello, author here. Most prompts include personalizing with a user's data. These prompts grow, and there is quite a bit of extra content that can be removed. U

Hello, author here. Most prompts include personalizing with a user's data. These prompts grow, and there is quite a bit of extra content that can be removed. UUIDs, pretty-printed JSON, ISO timestamps, repeated instruction blocks, etc., and all of it tokenizes to cost a lot. tokensift is a linter for that. It tokenizes the prompt with a real encoder the specific model uses: exact BPE counts for OpenAI and a calibrated estimate for Claude with the measured error shown. Then, it runs a set of rules against it and reports where the tokens are going, along with the dollar cost per finding. Works as a library or a CLI, and runs in CI. Current state: 20 rules, OpenAI and Claude supported, no Gemini yet.