Token Saving Mastery

节省 Token 的实战技巧 — 优化 AI 对话成本，提升性价比

Security Scan

The skill's advice on reducing tokens is plausible, but its runtime instructions tell the agent to list and patch other skills and create cron jobs and to export sessions externally — actions that exceed a simple helper and require elevated permissions and caution.

Purpose & Capability

Name/description claim token-saving tips — that matches most instructions. However, the runtime steps explicitly instruct listing all skills and patching other skills' SKILL.md to delete content (via skills_list, skill_manage patch). Editing other skills is not an obvious or low-risk requirement for a 'token-saving tips' helper and implies rights to modify unrelated artifacts.

Instruction Scope

SKILL.md tells the agent to call platform operations: session_search, skills_list, skill_view, skill_manage patch, hermes cron create, and to export sessions to external services (Obsidian/Notion). These instructions go beyond passive guidance — they direct programmatic modification of other skills and creation of background cron jobs, and recommend moving user conversation data to external endpoints. That broad scope could lead to unintended data exposure or destructive edits.

Install Mechanism

Instruction-only skill with no install spec or code files — nothing is written to disk by the skill itself. This is the lowest install risk.

Credentials

The skill requires no declared env vars or credentials, yet recommends exporting sessions to external services (Notion) and creating cron jobs that may need tokens or connectors. Also it presumes the ability to view and patch other skills and access session histories. The required permissions implied by the instructions are not declared or justified.

Persistence & Privilege

While always:false and no automatic install, the instructions explicitly instruct creating persistent cron tasks (hermes cron create) and editing other skills' prompts — both introduce persistent, system-wide effects. That increases blast radius if the agent is allowed to act autonomously or if cron prompts contain sensitive data.

What to consider before installing

This skill contains reasonable token-saving techniques, but it also instructs the agent to list and patch other skills and to create cron jobs and export sessions externally. Before installing or running it: 1) Verify what platform operations (skill editing, session access, cron creation) the agent is allowed to perform and restrict them if possible. 2) Don't allow automatic or background creation of cron jobs without review. 3) Require backups and code review before any skill_manage patch operations — editing other skills can break behavior or remove important content. 4) Avoid providing external-service tokens (Notion, etc.) unless you trust the workflow and audit exports. 5) Test in a sandbox account first and check activity logs for any edits or exports. If you cannot confirm who authored the skill or cannot limit its permissions, treat it cautiously or decline to install.

Token 节省大师 | Token-Saving Mastery

通过 4 大策略显著降低 AI 对话 Token 消耗，节省 30-70% 成本。

技能列表

1. 用 session_search 控制上下文 | session_search Context Control

难度： ⭐ 低

原理： AI 历史越长，消耗越多。只加载相关的记忆片段而非全部历史。

操作步骤：

1. 对话开始时，先调用 session_search 搜索关键词
2. 找到相关上下文后，整理成摘要注入对话
3. 让 AI 基于摘要继续，而非完整历史

适用场景： 多项目并行、长期开发、跨 session 协作

---

2. 减少技能文件体积 | Trim Skill File Bloat

难度： ⭐⭐ 中

原理： 每个 skill 的 system prompt 都会被加载，太大浪费。

操作步骤：

1. 用 skills_list 列出所有技能
2. 检查每个 skill 的 SKILL.md 体积
3. 删除以下内容：
   - 冗余的示例代码（保留 1-2 个即可）
   - 过时的 API 版本说明
   - 重复的步骤描述
   - 不常用的参考资料
4. 用 skill_view(name) 检查实际使用的行数
5. 用 skill_manage patch 精简

参考体积： 一个精简的 skill 应在 2-5KB 之间

---

3. 定期清理 session | Periodic Session Cleanup

难度： ⭐ 低

原理： session_history 无限累计，每次新建都会加载。

操作步骤：

1. 每周/每月运行一次 session 清理
2. 用 session_search 搜索关键项目名
3. 导出重要对话到外部笔记（Obsidian/Notion）
4. 标记不重要 session 为归档
5. 设置 cron 任务自动归档超过 30 天的 session

Cron 示例：

hermes cron create \
  --name "Session Cleanup" \
  --prompt "归档 30 天前的 session，保留最近 10 个" \
  --schedule "0 2 * * 0" \
  --skills "session_search"

---

4. Cron 任务用轻量 prompt | Lightweight Cron Prompts

难度： ⭐ 低

原理： Cron 任务在后台运行，不需要花哨的引导语和示例。

操作步骤：

1. 写 cron prompt 时遵守：
   - 直接说任务，不写"你是一个专业的..."
   - 不要示例，一个都不要
   - 结论导向："输出 X，返回 Y"
   - 限制输出长度："不超过 100 字"
2. 关联 skills 而非把知识写进 prompt
3. 避免多轮对话，用单次完成的任务

对比：

❌ 浪费写法：

你是一个专业的数据分析师，请帮我分析销售数据...
（200字引导 + 3个示例 + 200字总结要求）

✅ 节省写法：

分析附件 sales.csv，按地区统计月销售额，输出 TOP3 地区。

---

效果预估

策略	节省比例	实施难度
session_search 控制	20-40%	⭐
精简技能文件	10-20%	⭐⭐
定期清理 session	15-30%	⭐
轻量 cron prompt	10-25%	⭐

综合节省：30-70%

---

变现路径

1. ClawHub 技能市场 — 上传本技能，设置付费下载
2. 定制优化服务 — 按小时收费帮用户优化 Token 使用
3. 企业内部培训 — 打包成教程卖给团队
4. API 成本顾问 — 帮开发者分析并优化 AI 调用成本

---

验证步骤

完成设置后，运行：

hermes session list | wc -l  # 查看 session 数量
hermes skills list | wc -l   # 查看技能数量

对比优化前后的 API 调用账单，验证节省效果。