Analysis updated 2026-05-18
Lower your monthly AI coding assistant bill without changing how you work.
Cut filler language from AI chat replies while keeping code and tests precise.
Install a token-saving skill across Claude Code, Codex, Cursor, Windsurf, or Cline at once.
Switch between normal, lite, ultra, and off modes depending on the task.
| kulaxyz/token-diet | kelseyhightower/consul-on-kubernetes | kelseyhightower/serverless-vault-with-cloud-run | |
|---|---|---|---|
| Stars | 519 | 598 | 408 |
| Language | Shell | Shell | Shell |
| Last pushed | — | 2018-11-04 | 2022-04-28 |
| Maintenance | — | Dormant | Dormant |
| Setup difficulty | easy | hard | hard |
| Complexity | 1/5 | 4/5 | 4/5 |
| Audience | developer | ops devops | ops devops |
Figures from each repo's GitHub metadata at analysis time.
Installed with a single curl-to-bash one-liner, no separate configuration required.
token-diet is a small tool that makes AI coding assistants like Claude Code, Codex, Cursor, Windsurf, and Cline write less while still doing the same job correctly. AI coding tools are billed by the number of words, or tokens, they read and write, so cutting unnecessary text directly lowers the cost of using them, and the author reports an average bill reduction of around 31 percent across real test sessions, with no drop in correctness. It works by installing an always-on instruction set that tells the AI assistant to skip filler language, such as opening with Sure, here's the answer, or closing with let me know if you need anything else, and instead lead straight with the answer. The same philosophy applies across the whole session: documentation and comments get trimmed to only what is needed, tests focus on the important and risky cases instead of an exhaustive list, code is written to match exactly what was asked without extra unused pieces, and the assistant is encouraged to read only the specific lines it needs from a file rather than the whole thing. Correctness, full error messages, and exact commands are explicitly protected from being trimmed, so the assistant should never sacrifice accuracy just to save words. There are a few levels you can choose from: the full default behavior, a lighter mode that only trims conversational replies, a very terse mode for chat and progress updates while keeping code and tests precise, or turning it off entirely. Installing it is a single command that downloads and runs a script, which detects which coding assistant you have installed and sets things up automatically, either for your whole system or for a single project. Because these tools normally only load extra instructions when asked, the installer works around that by placing its instructions somewhere each supported tool always reads at the start of a session.
A skill that makes AI coding assistants like Claude Code and Cursor write less filler, cutting token costs by about 31% on average with no loss of correctness.
Mainly Shell. The stack also includes Shell, Claude Code, Codex.
Setup difficulty is rated easy, with roughly 5min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
Verify against the repo before relying on details.