一个极简的三层 LLM 技能系统,灵感来自 Karpathy 的 LLM-Wiki 理念
概览
一个极简的三层级大型语言模型(LLM)技能系统,灵感来自卡帕西的理念: 三个元技能作用于技能系统本身: 端到端流程(执行→蒸馏→指南,附有评测-公关攻略示例): install.sh 会运行布局检查并调用脚本/validate.py。 1. — 代理人阅读第2节 AGENT.md,路由到≤3技能,执行并收集蒸馏候选。2. : 3.——说“蒸馏这个”,代理会在skills/distill/SKILL.md中执行四阶段标准操作程序(SOP)。 - AGENT.md 纯路由+合同。- — 执行(技能只读)/ 蒸发(技能写入)/ 指南(仅写 AGENT.md)。- — 绝不对技能/进行批量读取;先匹配触发器,加载≤3 SKILL.md。- 每个技能——版本和状态。
概述(AI 翻译)
英语 |中文
一个极简的三层技能系统,针对LLM代理,灵感来自Karpathy的LLM-Wiki理念:
闭环
三个元技能作用于技能系统本身:
| 元技能 | 责任 | 输出 |
|---|---|---|
| ** 执行** | 读取AGENT.md路由→加载最小技能子集→完成任务→记录可重复使用信号 |
可交付 + distill-candidates |
| 蒸馏 | 精炼追踪到每个合同的新技能/更新技能 | skills/ /SKILL.md的变更 |
| ** 指南** | 维护AGENT.md路由表和技能健康 |
AGENT.md更改 |
仓库布局
端到端流程(执行→蒸馏 →指南,附有review-go-pr攻略说明):
安装
选项A — 独立仓库
选项B——安装到现有项目中
install.sh 运行布局检查并调用 scripts/validate.py。
快速开始
-
执行任务 — 代理阅读
AGENT.md§2,路由到≤3项技能,执行并收集蒸馏候选。 -
添加技能:
-
蒸馏陷阱 — 说“蒸馏这个”,代理人会在
skills/distill/SKILL.md中走进四阶段标准作业程序。
设计不变量
本页由 AI 翻译,技术细节请以原文为准。
README
以下为原文 README 节选,完整内容请查看原文。
llm-skill — Execute / Distill / Guide
English | 中文
A minimal three-layer skill system for LLM agents, inspired by Karpathy’s LLM-Wiki idea:
The closed loop
Three meta-skills operate on the skill system itself:
| Meta-skill | Responsibility | Output |
|---|---|---|
| execute | Read AGENT.md routing → load the smallest skill subset → finish the task → log reusable signals |
deliverable + distill-candidates |
| distill | Refine traces into new/updated skills per contract | skills//SKILL.md changes |
| guide | Maintain the AGENT.md routing table & skill health |
AGENT.md changes |
[Execute] ──traces──► [Distill] ──skill CRUD──► [Guide] ──routing──► [Execute] ...
Repository layout
llm-skill/
├── README.md # this file (English default)
├── README_zh.md # Chinese
├── CHANGELOG.md # append-only
├── AGENT.md # Schema: routing + contract (single source of truth)
├── AGENTS.md # Codex entrypoint
├── CLAUDE.md # Claude Code entrypoint
├── HERMES.md # HERMES entrypoint
├── install.sh # one-shot install / sanity check
├── scripts/
│ └── validate.py # front-matter + routing consistency checker
└── skills/
├── execute/ SKILL.md + SKILL_zh.md
├── distill/ SKILL.md + SKILL_zh.md
├── guide/ SKILL.md + SKILL_zh.md
└── _template/ SKILL.md + SKILL_zh.md
End-to-end pipeline (Execute → Distill → Guide, illustrated with a review-go-pr walkthrough):
Install
Option A — standalone repo
git clone llm-skill
cd llm-skill
bash install.sh
Option B — mount into an existing project
git clone .llm-skill
ln -s .llm-skill/AGENT.md ./AGENT.md
ln -s .llm-skill/AGENTS.md ./AGENTS.md # for Codex
ln -s .llm-skill/CLAUDE.md ./CLAUDE.md # for Claude Code
ln -s .llm-skill/HERMES.md ./HERMES.md # for HERMES
ln -s .llm-skill/skills ./skills
bash .llm-skill/install.sh
install.sh runs layout checks and calls scripts/validate.py.
Quick start
- Run a task — the agent reads
AGENT.md§2, routes to ≤ 3 skills, executes, and collects distill candidates. - Add a skill:
cp -r skills/_template skills/ # edit SKILL.md (and SKILL_zh.md if you want) # ask the agent to run `guide` to register it in AGENT.md §2.2 - Distill a pitfall — say “distill this” and the agent walks the four-phase SOP in
skills/distill/SKILL.md.
Design invariants
- Domain knowledge belongs only in
SKILL.md.AGENT.mdis pure routing + contract.
推荐工具
换一个关键词,或者移除筛选条件。
安装
npx skillfish add hanyuancheung/llm-skill