技能生态:新的工作模式 Skills & Ecosystem: New Ways of Working
会提问只是入门。真正的高效来自把经验沉淀成 Skill、插件、MCP 工具与 SOP——未必要写代码,把流程写清楚同样价值巨大。 Asking questions is just the start. Real leverage comes from distilling experience into Skills, plugins, MCP servers and SOPs — you don't have to write code; documenting a clear process is just as valuable.
能力的阶梯 The Ladder of Capability
从"用一次"到"处处可用",AI 工作模式逐级上升:
From "use once" to "always available", AI working modes climb step by step:
Skill 生态:把经验写进文件 Skill Ecosystem: Experience as Files
Skill(技能)是把"如何完成某类任务"的流程、规范、检查清单与示例,封装成一个结构化文件(例如 Claude 生态的 SKILL.md)。Agent 遇到对应任务时按需加载技能,就像给新人一本 SOP 手册。
A Skill packages the procedure, rules, checklists and examples for a task type into a structured file (e.g. SKILL.md in the Claude ecosystem). The agent loads it on demand — like handing a newcomer an SOP manual.
- 对你意味着什么:同一个技能,今天给 Qoder 用、明天给 Claude Code 用——沉淀一次,处处复用 Write once, use everywhere — today with Qoder, tomorrow with Claude Code
- 社区生态:大量开源技能库与"技能市场"出现(awesome 列表、插件市场)Open skill libraries and marketplaces are springing up
- 射电天文例子:把"FITS 立方体 → 矩图"的标准化流程做成技能,整个课题组共享 Turn your "cube → moments" pipeline into a shared group skill
一个技能文件长什么样 What a skill file looks like
# 技能:射电立方体矩图 · radio-moments
## 适用场景
输入 HI 21cm 谱线 FITS 立方体,输出 moment 0/1/2 图与 FITS。
## 步骤
1. 用 astropy 打印 header(CTYPE/CUNIT/CDELT/CRVAL/BUNIT/RESTFRQ),
确认轴顺序与速度单位(m/s 还是 km/s、LSRK 还是 barycentric)。
2. 用 spectral-cube 加载,统一速度为 km/s(radio 约定)。
3. 用 MAD 估计噪声,构建 3σ 掩膜,排除 NaN。
4. 计算 moment 0/1/2,保留 WCS 写出 FITS。
5. 打印各矩图峰值与单位;与 CASA immoments 交叉验证。
## 红线
- 不跳过第 1 步直接处理;
- 单位换算必须给出公式依据;
- 数值结论需可追溯。Plugin / 扩展生态 Plugin & Extension Ecosystem
IDE 与浏览器扩展 IDE & browser extensions
VS Code 扩展、浏览器插件:把 AI 能力织进你已有的工作流,而不是换个新工具。
VS Code extensions, browser plugins: weave AI into your existing workflow instead of switching tools.
Agent 插件市场 Agent plugin marketplaces
Claude Code 等已出现插件市场/插件协议,可安装第三方能力包;DeepSeek Harness 的理念就是 "Everything is a Plugin"(一切皆插件)——插件化让运行时轻装上阵、按需扩展。
Claude Code and others now ship plugin marketplaces/protocols; DeepSeek Harness's philosophy is "Everything is a Plugin" — a lean runtime that extends on demand.
MCP 生态:工具的"USB-C" MCP Ecosystem: USB-C for Tools
MCP(Model Context Protocol)把"模型 ↔ 工具"的对接标准化:写一次 MCP server,任何兼容 Harness/Agent 都能即插即用。目前已覆盖:
MCP standardizes model↔tool wiring: build an MCP server once, any compatible harness/agent can plug in. Coverage today includes:
- 文件系统、数据库、GitHub、浏览器、Slack/钉钉…… filesystem, databases, GitHub, browsers, Slack/DingTalk…
- 科学计算方向:社区已出现把数值库、数据归档封装为 MCP server 的做法 In scientific computing, folks are wrapping numeric libraries and data archives as MCP servers
- 射电构想:把 CASA 常用任务(导入、校准、成像)封装成 MCP server,Agent 就能"直接调用 CASA"而无需自己写脚本 Radio idea: wrap common CASA tasks (import, calibrate, image) as an MCP server so agents can "call CASA directly"
mcp_servers: {"casa": {"command": "casa-mcp"}},无需自己编程。
Many tools support declarative config to attach existing MCP servers — often just one line in a config file, no coding needed.
工作流与 SOP:未必要写代码 Workflows & SOPs: Code Is Optional
给 AI"编程"有三种层次,写流程文档同样算编程:
There are three levels of "programming" an AI — writing process docs counts too:
写代码 Write code
实现具体算法、管道、界面——传统编程。
Algorithms, pipelines, UIs — classic programming.
写提示词 Write prompts
把任务目标、约束、格式讲清楚——对 Agent 的"需求文档"。
Task goals, constraints, output formats — the "requirements doc" for agents.
写流程 / SOP Write processes / SOPs
把"先做什么、检查什么、红线在哪"写成 runbook/SOP——Agent 可照单执行,人也看得懂。
Steps, checkpoints, red lines as runbook/SOP — executable by agents, readable by humans.
一个 Runbook 示例(数据处理 SOP) A runbook example (data-processing SOP)
# Runbook:观测数据处理(示例)
## 输入
- 数据:/data/obs1.fits(或 MS 目录)
- 参考:CASA 官方指南
## 流程
1. 拷贝数据 → 校验 md5;
2. CASA 导入(importfits / importuvfits);
3. 校准:flagdata → setjy → gaincal → applycal;
4. 成像:tclean(说明参数依据);
5. 科学量:immoments / imfit;
6. 出图并归档:/output/ + README.md。
## 检查点
- 每步结束打印统计,与预期对比;
- 关键结果与独立软件/文献交叉验证;
- 数据与产物哈希归档,可复现。关键点:好的 SOP 文档本身就是"技能"——既能让人类新人照着做,也能让 Agent 照着执行,还能沉淀进 Skill/插件生态。DeepSeek Harness 的 goal(目标追踪)、workflow(工作流脚本)、subagent(子代理)正是这类机制的运行时实现。
Key idea: a good SOP is itself a skill — humans can follow it, agents can execute it, and it can be packaged into the Skill/plugin ecosystem. DeepSeek Harness's goal tracking, workflow scripts, and subagents are runtime implementations of exactly this.