360 Secure SkillHub
360 安全技能中心

Context Slimmer

该 Skill 审计始终加载到每条消息的工作区上下文文件(AGENTS.md、TOOLS.md、USER.md、MEMORY.md、HEARTBEAT.md、SOUL.md、IDENTITY.md),通过 scripts/measure.sh 测量当前 token 成本,按移动/删除/压缩三类给出精简建议和精简目标大小,最后输出各文件当前大小、预计大小与节省量报告。

sundevilatb
sundevilatb
热度 91.73
最近更新 2026-05-17 11:47:59
分类:效率依赖:代码执行 / 读写文件应用场景:报告撰写 / 摘要总结

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add context-slimmer
Metadata
name
context-slimmer
description
Audit and slim down always-loaded context files (AGENTS.md, TOOLS.md, USER.md, MEMORY.md, HEARTBEAT.md, SOUL.md, IDENTITY.md). Use when asked to reduce token usage, audit context files, optimize context window, or slim down workspace files. Measures current token cost and identifies what to move, remove, or compress.

Context Slimmer

Audit workspace files that load into every message and reduce their token footprint.

Quick Start

# Measure current context cost
bash scripts/measure.sh

# Full audit (outputs recommendations)
bash scripts/measure.sh --audit

Audit Process

For each always-loaded file, evaluate:

  1. Move to skill — Content only needed for specific tasks (betting config, group chat rules, detailed protocols). Move to a skill or reference file that loads on demand.
  2. Remove — Outdated info, dead features, completed one-time setup, duplicated content across files.
  3. Compress — Verbose explanations that could be 1 sentence. If the agent already knows it, cut it.

Rules of Thumb

  • If a cron job handles it, remove it from HEARTBEAT.md
  • If it's in SOUL.md, don't repeat it in MEMORY.md or AGENTS.md
  • If it's in USER.md, don't repeat it in MEMORY.md
  • If the agent does it daily, it doesn't need instructions — just a trigger word
  • Prefer 1 sentence over 5 bullets saying the same thing
  • Target: each file should justify every line's token cost

Expected File Sizes (lean targets)

File Target
AGENTS.md < 500 tokens
TOOLS.md < 500 tokens
USER.md < 700 tokens
MEMORY.md < 400 tokens
HEARTBEAT.md < 400 tokens
SOUL.md < 250 tokens
IDENTITY.md < 50 tokens
Total < 2,800 tokens

Output Format

Report: current size, projected size, savings per file. Include specific recommendations grouped by move/remove/compress.