token-diet

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

token-diet

Always-on token-efficiency skill for coding agents — Claude Code, Codex, Cursor, Windsurf, Cline. Trims tokens across the whole session (replies, docs, tests, code, context, tool use) without losing correctness. ≈31% lower bill on average (−17% to −54% by session type) and −30% to −81% output on real Sonnet 5 runs.

trim the fat, keep the muscle.

Install

# one-liner — auto-detects Claude Code, Codex, etc.:
curl -fsSL https://raw.githubusercontent.com/Kulaxyz/token-diet/main/install.sh | bash

# pass options through `bash -s --`, e.g. the telegraphic level:
curl -fsSL https://raw.githubusercontent.com/Kulaxyz/token-diet/main/install.sh | bash -s -- --ultra

Or clone and run ./install.sh (--ultra telegraphic chat · --uninstall remove). Target one agent with -a claude|codex|cursor|windsurf|cline|all|print, or add --project to install into the current repo instead of globally. It runs always-on — no per-message command — but /token-diet [on|lite|ultra|off] also works on demand.

What it does

• Replies — lead with the answer; no preamble (“Sure! Here’s…”) or postamble (“Let me know…”); report deltas, not narration.
• Docs / memory / hand-offs / plans / comments — minimum words that still say everything; comment the non-obvious why, not the code.
• Tests — only key + critical/edge paths, grouped; ≤10 per session; never skip money/auth/data-loss.
• Code — build only what’s asked (YAGNI); concise but idiomatic; no dead code; never cryptic.
• Context — grep before you read; read only the lines you need, never whole files; batch reads; fewer turns; never re-read a just-edited file.
• Tools — batch independent calls; stop the moment you have enough to act; targeted test runs while iterating.
• Sub-agents — delegate broad bounded search to a cheaper model; keep correctness-sensitive verification yourself; tight, complete instructions.
• Guardrails — concision applies to output, never to reasoning: correctness, critical test coverage, and verbatim code/commands/errors are off-limits.

Levels

on (default — everything above) · lite (communication + artifacts only) · ultra (telegraphic chat/progress; code, tests, docs stay precise) · off.

Before / After — ultra chat reply (token counts measured, o200k_base)

Normal — 102 tokens: “Great question! The reason your Stripe webhook signature verification is failing is that you’re passing the already-parsed JSON body into constructEvent. Stripe computes the signature over the exact raw bytes of the request, so once your framework has parsed the body into JSON the bytes no longer match and verification fails. What you want to do is read the raw request body instead — for example with req.text() in a Next.js route handler — and pass that string to constructEvent. Let me know if you’d like a code snippet!”

token-diet ultra — 34 tokens (−66%): “constructEvent verifies over raw bytes → parsed JSON fails. Read raw body (req.text() / Buffer), pass that to constructEvent. Fix at webhook/route.ts.”

Same fix, same identifiers, same path — only filler gone.

Numbers (real Sonnet 5 runs)

Output savings are consistent; the bill win depends on how much of the session is output vs. unavoidable file-reading:

Session type	Output	Bill
Output-heavy (advice, planning, explanation)	−81%	−54%
Code change + tests (nestjs/nest, 1673 files)	−49%	−22%
Read-heavy comprehension	−30%	−17%
Average	−53%	−31%

Average is unweighted across the three scenarios; the 54% is best-case, not typical. Correctness held in every run. Method + full tables: bench/RESULTS.md; reproduce with ANTHROPIC_API_KEY=… node bench/bench.mjs.

How always-on works

A skill body loads on-demand, so to make it always apply the installer injects the directive through a channel that fires every session: a SessionStart hook (Claude Code), or an always-loaded context file (Codex AGENTS.md, Cursor/Windsurf/Cline rule files). Same trick caveman uses.