Asif

Token Optimizer for OpenClaw & Hermes

AI 智慧體v4.1.1
作者:Asif更新於 2026-09-19共 25 個版本 · 最新版本釋出於 2026-09-19openclaw-token-optimizer@clawhub_asif2bd/openclaw-token-optimizer
工作流編排上下文管理工具呼叫#Document#Json

據 AlphaMoat 統計,截至 2026-10,Token Optimizer for OpenClaw & Hermes 累計下載 28,344 次、獲 30 個星標,按累計下載量在本次參與排名的 10,914 個 AI 技能中排名第 989 位,在「AI 智慧體」分類(1,632 個)中排名第 150 位。

列表預設按累計下載量排序,也可按星標排序;詳情頁排名按累計下載量計算。

累計下載
2.8萬
星標
30
全站下載量排名
#989
全站前 10% · 「AI 智慧體」前 10%
平台累計下載星標評論
ClawHub ↗1.4萬302
SkillHub ↗2.8萬30—

各平台的計數口徑不同,統計時間不一致。

這個技能是做什麼的

檢查模型選擇、較大的指令檔案和定期任務,找出可能的 token 使用低效處並提出後續建議,不會自行修改設定。來源說明,節省空間只是待驗證的機會,並非已測得的帳單降幅。技能不會列印原始設定或環境變數值,但選取的設定與 profile 檔案仍可能含有敏感資訊。OpenClaw 原生診斷讀取可能連線至已設定的服務或更新本機快取。產生的報告可能揭露模型名稱、agent ID、檔案名稱或運作情況,分享前應先檢查。

資料釋出於 2026-10-27

可以用來做什麼

  • 找出可能造成 token 浪費、值得複核的定時任務或模型選擇。
  • 找出可能讓上下文變大的指令檔案。

什麼時候用

  • 想排查 AI 使用中可能的低效問題時。
  • 需要複核模型選擇或定期任務時。

示例提示詞

以下是根據技能文件生成的示例,不代表實際執行結果。

檢查我的定期任務中是否有可能不必要的 AI 呼叫。
哪些指令檔案可能讓上下文變得很大?

月度資料趨勢

累計下載2.8萬
95162萬3萬2026-03 · 累計下載 1.1萬2026-05 · 累計下載 1.3萬2026-06 · 累計下載 1.5萬2026-07 · 累計下載 1.6萬2026-08 · 累計下載 1.6萬2026-09 · 累計下載 1.8萬2026-10 · 累計下載 2.8萬26-0326-0926-10
星標30
2829302026-03 · 星標 282026-05 · 星標 302026-06 · 星標 302026-07 · 星標 302026-08 · 星標 302026-09 · 星標 302026-10 · 星標 3026-0326-0926-10

2026-09 起統計方法調整:同名技能按作者分別統計,並覆蓋全部技能。圖中虛線為調整前的月份,與之後的資料不直接比較,也不計算跨越調整的增長。

月份累計下載環比星標
2026-102.8萬—30
2026-091.8萬—30
2026-081.6萬+3.3%30
2026-071.6萬+6.0%30
2026-061.5萬+13.3%30
2026-051.3萬—30
2026-031.1萬—28

SKILL.md 原文

以下是作者釋出的原始檔案,按純文字顯示,本站未做修改或稽核;其中的連結不可點選,命令請核實後再執行。抓取於 2026-10-23。

5.4 KB · ClawHub v4.1.1
---
name: openclaw-token-optimizer
description: "Find potential token waste in OpenClaw or Hermes Agent. Use when asked to reduce AI costs, review model choices, check large instructions or audit recurring tasks. Explain what matters and suggest safe next steps without changing settings."
version: 4.1.1
author: Asif2BD
homepage: https://missiondeck.ai
source: https://github.com/Asif2BD/OpenClaw-Token-Optimizer
license: MIT
openclaw: ">=2026.9.4"
metadata: {"openclaw":{"emoji":"💰","requires":{"bins":["python3"]}}}
展開全部(共 108 行)
---

# Make your AI agent leaner—not less capable

Ask your agent:

> “Check where I might be wasting tokens and tell me what to improve.”

Or try:

- “Review my scheduled tasks for unnecessary AI usage.”
- “Which instruction files are making my context heavy?”
- “Help me choose a suitable model for routine work.”

Works with **OpenClaw and Hermes Agent**. Your agent handles the technical steps below;
you do not need to copy Python commands or prepare exports when the required evidence
is already available in its environment.

## What you get

A short, prioritized explanation of what was checked, what deserves attention and what
to do next. Nothing silently changes your models, instructions or schedules. Potential
savings are opportunities to test—not a promise or a measured reduction in your bill.

Built by [MissionDeck.ai](https://missiondeck.ai).

## Instructions for the agent

### 1. Start with what you already know

When the user requests an audit, proceed with relevant read-only checks; do not ask them
to restate the runtime, agent ID or active profile if trusted session context already
provides it. Say briefly that you will inspect and recommend, without changing settings.

Use this decision order:

- An explicit user-selected runtime/profile takes priority.
- Otherwise use the hosting session's known runtime and active agent/profile.
- Installed binaries alone are not proof of which profile is active. If the host is unknown
  or multiple profiles are plausible, ask one focused question rather than auditing the
  wrong system. Continue any independent checks whose scope is already clear.

Resolve bundled script paths relative to this SKILL.md directory. Use an available Python
3.10+ interpreter. No automatic installs, recursive home-directory searches or credential
searches. Never print raw configuration or environment values.

### 2. Run the matching read-only audit

**In OpenClaw:** use the current agent ID from trusted runtime context, not the example
name or the CLI's default. Follow `references/QUICKSTART.md` for the command. Include
`--config` only when the active config path is already established and the audit needs it.
If the CLI is unavailable, use exports already supplied; otherwise report the missing
capability instead of asking the user to run a long list of commands.

**In Hermes:** use the active profile path already supplied by the host/user. A known
HERMES_HOME path is profile evidence; do not dump the environment or load `.env`. Do not
assume the default profile just because a default directory exists. The audit reads only
config.yaml and cron/jobs.json. Prefer the existing Hermes Python environment for YAML
support. If no cron file exists, report automation coverage as unavailable; audit the
available config separately rather than claiming zero jobs. See references/QUICKSTART.md.

Do not ask for catalog/status exports for an ordinary Hermes profile audit: those are
only needed for the separate model-routing function. Do not start Hermes or any job.

### 3. Match extra checks to the question

- Context: inspect only known, relevant instruction files. Character/4 estimates are not
  actual injected tokens; use native context diagnostics if available. Never delete instructions.
- Model choice: use the user's policy and known catalog/status evidence. Missing capability
  metadata means unverified, not unsupported. No automatic switching or paid canary.
- Spending: without complete usage evidence, say actual spending/savings are unknown.
  Native usage ingestion and automatic budget alerts are not implemented.

Read only the reference needed: QUICKSTART.md for commands, HERMES.md for profile/model
precedence, PROVIDERS.md for routing/cache advice. Do not load every reference by default.

### 4. Explain the result, not the machinery

Lead with a plain-language verdict and at most three prioritized next steps. For each,
explain the evidence, why it matters, and the uncertainty. Separate confirmed problems
from review candidates; do not count every informational finding as a fault. End with
what was not checked and confirmation that no settings were changed.

Do not paste raw JSON unless requested. Offer technical details only when useful. If
nothing actionable is found, say so. Never claim a fixed savings percentage, cheapest
model, successful authentication or measured dollars without evidence.

## Boundaries

All bundled commands are diagnostic. No config edits, installs, restarts, messages, new
jobs or credential reads. OpenClaw native reads may contact configured services and update
native caches. Hermes support is file-based, not native catalog/authentication discovery;
profile overlays and runtime overrides remain unverified. See SECURITY.md.

Python 3.10+; optional PyYAML for YAML input. OpenClaw adapter tested on 2026.9.4; Hermes
file compatibility is pinned in references/HERMES.md. The CLI still requires explicit
runtime selection for Hermes; the agent chooses it from trusted context, not autodetection.

「AI 智慧體」分類下的熱門技能

Asif 的其他技能

資料月份:2026-10 · 下載/星標數聚合自公開技能註冊庫(ClawHub、SkillHub),按月更新。口徑說明見資料方法。