Context Brief · 2026-08-22
MCP answers “what can the agent do.” Skills answer “how should it think.” The working pattern is Skill orchestration plus MCP execution. Pick one only when the hard part is only access, or only judgment.
These are not two products you rank. MCP is an open protocol for capability access. Agent Skills are a folder format for packaged judgment. One is a USB-C port. The other is a runbook. Teams waste weeks when they build a server for a workflow problem, or a markdown file for a live-data problem.
Model Context Protocol is an open standard for connecting an AI app to external systems. Tools, resources, and prompts travel over a shared client-server contract. Auth, transport, and schemas live on the server. The model sees a tool list and emits structured calls. Anthropic introduced it in November 2024 and donated it to the Linux Foundation’s Agentic AI Foundation in December 2025.
Agent Skills are a lightweight open format: a folder with a required SKILL.md, optional scripts, and references. The name and description load first. The body loads when the task matches. Scripts and files load only if needed. Anthropic shipped the pattern in Claude in October 2025 and published it as an open standard on 18 December 2025.
A grant of access. Query this API. Read this dataset. Write to this system. Shared across agents once the server exists.
Packaged expertise. When you see X, do Y first, then Z. Output looks like this. Load only when the task matches.
David Girvin at Sumo Logic asks one question: if a senior person did this by hand, is the hard part getting the data, or knowing what to do once they have it? Access → MCP. Judgment → Skill. Most of the time, if you are honest, the hard part is judgment.
You need live systems: APIs, databases, logs, clusters, CRMs. Several agents should share one capability. The pipeline must carry auth and permissions. The value is the data or the write, not the write-up.
You need a repeatable workflow with domain rules. You want a light, local instruction set. Output format must stay consistent. The team’s expertise is the asset, and it should version like a runbook.
Skill (reasoning layer) ├─ Receives user intent ├─ Decides what data and writes are needed ├─ Calls named MCP tools ├─ Applies domain rules to the result └─ Returns a structured artifact MCP (execution layer) ├─ Publishes tool schemas ├─ Handles auth and transport ├─ Runs the API or process call └─ Returns raw data
A skill can compose several MCP tools. The server stays centrally maintained, so a schema change lands once. The skill stays a versionable knowledge file. That is the split Red Hat describes for DevOps: MCP reaches the cluster and the ticket system; the skill holds the incident runbook.
| Signal | Choose | Why |
|---|---|---|
| Live data or a write to an external system | MCP | The model cannot invent current inventory, logs, or ticket state. |
| Several agents need the same capability | MCP | One server, many clients. Auth and schemas stay in one place. |
| The hard part is permission and transport | MCP | Standardized read/write beats a one-off wrapper in every agent. |
| Sequence, hypotheses, and “do Y before Z” | Skill | Access without a workflow is a new hire with every badge and no training. |
| Stable process, team voice, fixed output shape | Skill | Markdown is cheap to edit and easy to version like a runbook. |
| Docs or APIs change faster than your playbook | MCP first | LlamaIndex found a live docs server beat local skills that went stale. |
| Process is stable; knowledge is the scarce thing | Skill first | Capture judgment. Call MCP only for the live bits. |
| A real workflow that both decides and fetches | Both | Skill names the tools and the order. MCP does the calls. |
Rule of thumb from LlamaIndex: fast-changing single source of truth → MCP. Stable process you want to encode → Skills.
| Dimension | MCP | Skill |
|---|---|---|
| Setup | Auth, transports, schemas. Usually a developer job. | A folder and a Markdown file. Almost no install. |
| Execution | Deterministic call once the tool is chosen. | The model interprets instructions. The “how” can drift. |
| Latency | Network on every call. | Local load. No extra hop unless the skill then calls MCP. |
| Maintenance | Server update reaches every client. | You edit Markdown. Stale skills are a real failure mode. |
| Scale | Too many tools and discovery gets noisy. Gateways appear. | Easy to add. Selection quality drops as the catalog grows. |
| Context cost | Tool schemas can bloat the window. Fat responses make it worse. | Name and description stay small. Body loads on demand. |
| Failure mode | Wrong tool, or a tool that returns too much. | Misread instructions, then the wrong tool on top. |
LlamaIndex’s own coding-agent test is a warning, not a universal win for MCP: when the scarce asset is current SDK docs, a docs server beat skills they had to keep rewriting.
| Layer | Do this | Avoid this |
|---|---|---|
| MCP deployment | Streamable HTTP for cloud APIs (zero install, OAuth). Bundle local files and processes when the work is on the machine. | One giant “do everything” server that cannot be shared or audited. |
| Tool grain | Split by operational meaning: search logs, open a ticket, list deploys. | Tools so large they hide a workflow, or so small the agent drowns in names. |
| Errors | Auth failures, timeouts, and retries at the tool layer. | Leaking raw 401s into the skill and hoping the model improvises. |
| User UI | Elicitation for simple forms. MCP Apps when the UI is the product. | Pretending chat can collect a twelve-field change request. |
| Skill naming | Gerunds the agent can cite: analyzing-sales-data, debugging-incident. | Vendor slogans or names that do not say when to load. |
| Skill body | Triggers, step order, decision rules, output shape, and the MCP tools it may call. | A vibe paragraph with no stop conditions and no named tools. |
| Skill ops | Design, test, and version like a runbook. Let the agent load on demand. | Dumping every skill into the system prompt “just in case.” |
| Move | What it looks like |
|---|---|
| Discover before you bind | A skill such as build-mcp-server asks who the users are, what they must read or write, and how auth works, before you pick HTTP versus a local bundle. |
| Name the tools in the skill | List which MCP tools the skill may call, and in which case. Do not leave the model to invent a toolbox. |
| Keep three layers | Low: raw APIs on MCP. Mid: one business process in a skill. High: several skills for a cross-domain job. |
| Do not auto-publish | If a channel needs a human (login, tone, legal), the skill must stop and hand off. A browser MCP is not a license to send. |
| Job | Skill does | MCP does |
|---|---|---|
| Sales analysis | Clean → aggregate → visualize. Names the chart and the caveats. | Fetches live sales rows the model must not invent. |
| DevOps | deploying-service and debugging-incident hold order, thresholds, and the postmortem shape. | One server for CI, logs, and monitors. Many skills reuse it. |
| Cross-border commerce | Product pick → price → listing. Encodes the decision chain. | Talks to TikTok Shop and Amazon. Auth stays on the server. |
| This content pack | tao-content-brief decides angle, tables vs prose, and what not to claim. | Creates the project, moves the workspace, deploys the site. Publish stays human. |
| Trap | What happens |
|---|---|
| MCP for a workflow problem | The agent can query everything and still investigate nothing. Sumo’s SOC example: tools without CloudTrail-then-IAM-then-tags. |
| Skill for a live-data problem | The markdown goes stale the week the SDK changes. LlamaIndex hit this on their own docs. |
| Both, with no contract | The skill never names tools. The server never knows which workflow it serves. You get two catalogs and no orchestration. |
| Treat this page as a bake-off | There is no public score that holds the model and the task fixed. Pick by the hard part, not by stars. |
MCP intro · Agent Skills specification · Anthropic on Agent Skills · Sumo Logic, 4 Mar 2026 · LlamaIndex, Skills vs MCP tools · Red Hat Developer, 25 May 2026
Also in the newsletter.
One line: MCP is the badge. Skills are the training. Use both, and let the skill call the tools.
MCP 回答「代理能做什么」。Skills 回答「它该怎么想」。能跑通的结构是 Skill 编排、MCP 执行。只有当难点只在取数，或只在判断时，才单独选一边。
这不是两款可以排名的产品。MCP 是能力接入的开放协议。Agent Skills 是把判断装进文件夹的格式。一个像 USB-C，一个像 runbook。团队最常浪费的几周，是给流程问题写服务器，或给实时数据问题写 Markdown。
Model Context Protocol 是连接 AI 应用与外部系统的开放标准。工具、资源和提示走同一套客户端-服务器契约。鉴权、传输和 schema 在服务器上。模型看到工具列表，再发出结构化调用。Anthropic 于 2024 年 11 月提出，2025 年 12 月捐赠给 Linux 基金会的 Agentic AI Foundation。
Agent Skills 是轻量开放格式：一个目录，必须有 SKILL.md，可选脚本和参考资料。名称和描述先加载。正文在任务匹配时加载。脚本和文件按需再读。Anthropic 于 2025 年 10 月在 Claude 落地，2025-12-18 作为开放标准发布。
准入。查这个 API，读这批数据，写进这个系统。服务器一旦存在，多个代理可以共用。
封装过的经验。看到 X，先做 Y，再做 Z。输出长这样。只在任务匹配时加载。
Sumo Logic 的 David Girvin 只问一句：如果让团队里的老人空手做这件事，难点是拿到数据，还是拿到之后知道怎么处理？接入难 → MCP。判断难 → Skill。多数时候，如果你诚实，难点是判断。
需要活系统：API、数据库、日志、集群、CRM。多个代理要共享同一能力。管道必须带鉴权和权限。价值在数据或写入，不在成文方式。
需要带领域规则的可重复流程。想要轻量、本地的说明书。输出格式必须稳定。团队经验才是资产，应该像 runbook 一样版本化。
Skill（推理层） ├─ 接收用户意图 ├─ 判断需要哪些数据和写入 ├─ 调用点名的 MCP 工具 ├─ 按领域规则处理结果 └─ 交出结构化产物 MCP（执行层） ├─ 公布工具 schema ├─ 处理鉴权与传输 ├─ 执行 API 或进程调用 └─ 返回原始数据
一个 Skill 可以组合多个 MCP 工具。服务器集中维护，schema 改一次就传到所有客户端。Skill 仍是可版本化的知识文件。这就是 Red Hat 写的 DevOps 拆法：MCP 连集群和工单；Skill 装着事故 runbook。
| 信号 | 选 | 原因 |
|---|---|---|
| 要实时数据，或要写入外部系统 | MCP | 模型不能编造当前库存、日志或工单状态。 |
| 多个代理需要同一能力 | MCP | 一台服务器、多个客户端。鉴权和 schema 只维护一份。 |
| 难点是权限和传输 | MCP | 标准化读写，好过每个代理各写一层包装。 |
| 顺序、假设、「先 Y 再 Z」 | Skill | 只有接入没有流程，等于新人拿齐所有工牌却没受过训。 |
| 流程稳定，团队语气和输出形状要固定 | Skill | Markdown 便宜，也好像 runbook 一样改版本。 |
| 文档或 API 比你们的手册变得更快 | 先 MCP | LlamaIndex 发现：活文档服务器，胜过他们不断改写的本地 Skill。 |
| 流程稳定，稀缺的是判断 | 先 Skill | 先把判断写下来。只为活数据调用 MCP。 |
| 既要决策，又要取数的真流程 | 两边一起 | Skill 点名工具和顺序。MCP 负责调用。 |
LlamaIndex 的经验法则：场景变得快、需要单一事实源 → MCP。流程稳定、要沉淀团队知识 → Skills。
| 维度 | MCP | Skill |
|---|---|---|
| 上手 | 鉴权、传输、schema。通常是开发工作。 | 一个目录加一份 Markdown。几乎不用装。 |
| 执行 | 工具一旦选对，调用是确定的。 | 模型解读指令。「怎么做」会漂。 |
| 延迟 | 每次调用都有网络。 | 本地加载。除非 Skill 接着调 MCP，否则没有额外跳。 |
| 维护 | 服务器一改，所有客户端都更新。 | 你改 Markdown。过期 Skill 是真实故障。 |
| 规模 | 工具一多，发现就吵。于是出现网关。 | 加起来容易。目录变大后，选对的难度上升。 |
| 上下文成本 | 工具 schema 会撑爆窗口。胖响应更糟。 | 名称和描述很小。正文按需加载。 |
| 失败形态 | 选错工具，或工具一次吐太多。 | 先读错指令，再叠一层选错工具。 |
LlamaIndex 自己的编码代理试验是警告，不是 MCP 的全面胜利：当稀缺资产是最新 SDK 文档时，文档服务器胜过他们得不断重写的 Skill。
| 层 | 这样做 | 别这样做 |
|---|---|---|
| MCP 部署 | 云 API 优先 Streamable HTTP（免安装、可 OAuth）。本机文件和进程再打成本地包。 | 一台「什么都能做」的巨型服务器，既无法共享也无法审计。 |
| 工具粒度 | 按操作语义拆：搜日志、开工单、列发布。 | 大到藏起整个流程，或小到代理被名字淹死。 |
| 错误 | 鉴权失败、超时、重试放在工具层统一处理。 | 把裸 401 丢给 Skill，指望模型临场发挥。 |
| 用户界面 | 简单表单用 Elicitation。界面本身是产品时用 MCP Apps。 | 假装聊天能收齐一张十二字段的变更单。 |
| Skill 命名 | 用代理能引用的动名词：analyzing-sales-data、debugging-incident。 | 厂商口号，或看不出何时该加载的名字。 |
| Skill 正文 | 触发条件、步骤顺序、决策规则、输出形状，以及它可调用的 MCP 工具。 | 一段气氛描写，没有停止条件，也不点名工具。 |
| Skill 运维 | 像 runbook 一样设计、测试、版本化。让代理按需加载。 | 「以防万一」把所有 Skill 塞进系统提示。 |
| 动作 | 长什么样 |
|---|---|
| 先问清楚再绑定 | 像 build-mcp-server 这样的 Skill，先问清用户是谁、要读还是写、鉴权怎么做，再选 HTTP 还是本地包。 |
| 在 Skill 里点名工具 | 写明这个 Skill 可以调哪些 MCP 工具、在什么情形下调用。不要让模型自己发明工具箱。 |
| 保持三层 | 底层：MCP 上的原始 API。中层：一个 Skill 封装一条业务过程。上层：多个 Skill 做跨域任务。 |
| 不要自动发布 | 渠道需要人（登录、语气、法务）时，Skill 必须停下来交接。浏览器 MCP 不是发送许可。 |
| 工作 | Skill 做 | MCP 做 |
|---|---|---|
| 销售分析 | 清洗 → 聚合 → 可视化。点名图表和限制条件。 | 拉取模型不能编造的实时销售行。 |
| DevOps | deploying-service 和 debugging-incident 管顺序、阈值和复盘结构。 | 一台服务器覆盖 CI、日志和监控。多个 Skill 复用。 |
| 跨境电商 | 选品 → 定价 → 上架。封装整条决策链。 | 对接 TikTok Shop 和 Amazon。鉴权留在服务器。 |
| 这套内容包 | tao-content-brief 决定角度、表格和正文怎么分、什么不能写。 | 建项目、切工作区、部署站点。发布仍由人粘贴。 |
| 陷阱 | 后果 |
|---|---|
| 用 MCP 解决流程问题 | 代理什么都能查，却仍然不会调查。Sumo 的 SOC 例子：有工具，却没有「先 CloudTrail，再 IAM，再看敏感标签」。 |
| 用 Skill 解决活数据问题 | SDK 一改，Markdown 就过期。LlamaIndex 在自己的文档上撞过。 |
| 两边都上，却没有契约 | Skill 不点名工具。服务器不知道服务哪条流程。最后是两份目录，没有编排。 |
| 把这页当成烘焙赛 | 没有公开分数能固定模型和任务。按难点选，不要按 star 选。 |
MCP 简介 · Agent Skills 规范 · Anthropic 谈 Agent Skills · Sumo Logic，2026-03-04 · LlamaIndex，Skills vs MCP tools · Red Hat Developer，2026-05-25
Also in the newsletter.
一句话：MCP 是工牌。Skills 是训练。两边都要，并且让 Skill 去调工具。
Leave your email and I’ll send new essays when they go up.