Skills, MCP, and plugins — the stack in plain words
Goal: One beginner table for plugins, skills, and MCP connectors—what each does, one example, main risk—then propose the thinnest stack.
Best for: Beginners who saw Skills / MCP / Plugins in release notes.
Add one layer at a time. Inventory what you installed.
Plugins, skills, and MCP are layers—not synonyms. Pick the thinnest layer that does the job; stack only when a weekly workflow needs all three.
Why beginners skip this
Turning on plugins, skills, and MCP together feels complete—debugging cannot tell who moved. Run one channel first.
Step 1 — Three layers
- Skill = the playbook (how to do the job).
- MCP = a way for the AI to use outside tools/data (mail, docs, calendars) through a standard door.
- Plugin = an installable bundle that can include skills and/or MCP connections — like an app pack.
Step 2 — Cooking metaphor
Keep the kitchen metaphor: do not mix “seasoning” with stove permissions.
| Layer | Kitchen metaphor |
|---|---|
| Skill | Recipe card |
| MCP tools | Stove, fridge, delivery app — real appliances |
| Plugin | Meal kit box with recipe + coupons to appliances |
You can use a recipe alone (skill in a Project). You only open appliances (MCP) when the job needs live data or actions — and you keep Ask-before-act on.
Step 3 — Beginner order
- 1) Write a clear prompt pattern (Goal / Format / Boundaries).
- 2) Save it as Project instructions or a skill draft.
- 3) Only then connect one safe, read-mostly MCP tool.
- 4) Install a plugin when you want a maintained bundle — still review permissions.
Example
Filled prompt
I have: 1 Custom GPT, 2 skills, 1 MCP Drive read.
Draw a one-page stack map: which job uses which layer. Flag overlap. Recommend one thing to uninstall this week.
What good output looks like
- One-page map: Custom GPT / skills / MCP Drive each tied to a named job
- Flags overlap where two layers claim the same weekly task
- Recommends one concrete uninstall or disconnect for this week
- Keeps MCP on read-mostly—no silent Gmail send added
- Beginner can act from the map without a protocol essay
Example
Filled prompt
Inventory GPT+2 skills+Drive MCP.
One-page map jobs→layer. Uninstall one overlap this week.
Context: real workstream this week (Asia/Shanghai). Name actual files/people.
Constraints: no invented metrics/owners/prices; label Unknowns; stop before send/delete/pay.
Done check: I can mark pass/fail in one minute on the output.
What good output looks like
- Proposes skill/Project instructions only for notes → bullets
- Calendar stays draft-in-chat; no auto-create events yet
- Explicit “do not connect Gmail send / Drive write this month”
- Stack fits on one sticky note—no three-year architecture diagram
- Ask-before-act called out if any connector is added later
Save this card
Card: Skills = playbooks · Plugins = install surface · MCP = channel. Name which layer owns send/delete before you stack.
Before you start
Open the listing and a notes doc. Score purpose/scopes/uninstall for “See how the three layers fit so new product” before Install. Prefer wait over Always allow. Guide: plugins-skills-mcp-stack. Keep the real job visible the whole time; a blank box means stop and fetch the artifact first.
How to use this week
This week pick one channel (plugin / skill / MCP) and finish one real job; then note why you did not pick the other two.
Thinnest layer wins
Prefer the smallest layer that does the weekly job. Stack only when needed.
One job, one layer
If you need “and” to describe why you installed something, you are stacking early. A skill that shapes notes is enough for many weeks; MCP calendar write is a later promotion with Ask-before-act still on.
Example
Filled prompt
Map: Custom GPT (FAQ) · skill (status) · MCP Drive read.
Find one overlap to uninstall this week. One-page stack map.
Context: real workstream this week (Asia/Shanghai). Name actual files/people.
Constraints: no invented metrics/owners/prices; label Unknowns; stop before send/delete/pay.
Done check: I can mark pass/fail in one minute on the output.
What good output looks like
- Eval score written before install for “See how the three layers fit so new product names ”
- Uninstall path and scopes listed
- No always-allow on write/send
- Pass/fail: keep-or-remove decision dated
Red flag
Unclear which layer owns send/delete; stacking all three with no job—draw the layers first.
Pitfalls
| Pitfall | Fix |
|---|---|
| Installing everything at once | One layer at a time |
| MCP write tools before skills hygiene | Skills/docs first; MCP least privilege |
| No inventory of what you installed | Keep a one-page stack map |
Example
Filled prompt
Temptation: skip the habit for plugins-skills-mcp-stack to save five minutes.
Job still: See how the three layers fit so new product
Smallest correct pass: Done visible · checkable artifact · Unknowns [?].
Boundaries: draft only; no silent send; no invented metrics.
Return: artifact + refuse-line said aloud.
What good output looks like
- Names the shortcut refused (specific to plugins-skills-mcp-stack)
- Still delivers checkable output for “See how the three layers fit so new product”
- Unknowns labeled; draft-only
- Pass/fail: refuse-line dated (plugins-skills-mcp-stack #4)
Practice now
Retirement line: name what you will uninstall if unused 30 days (Asia/Shanghai).
Cost-of-layer: estimate minutes each layer adds tonight (Asia/Shanghai).
Delete-one: remove one unused layer from the diagram (Asia/Shanghai).
Layer-order drill: decide plugin→skill→MCP order for one job; write why (Asia/Shanghai).
One-layer drill: solve the job with one layer only first (Asia/Shanghai).
Stack drill: draw plugin vs skill vs MCP for one job in three bullets (Asia/Shanghai).
Thin-spot drill: rewrite the Save card with real nouns from your work (repo, doc, inbox). Generic placeholders fail the drill.
Tonight enable only one layer (plugin or skill or MCP), run one real job through, then add the next layer.
Close the loop
Write one line: next time I turn on all three channels at once, I refuse to run.
One sentence takeaway
One channel first; then note why not the other two.
Next
Skill、MCP、插件——用大白话把这一叠看清
目标: 用一张新手表分清插件、skill、MCP 连接器:各干什么、各举一例、主要风险是什么;再提一个最薄的栈。
适合: 想一次上齐三层、却还没稳定聊天习惯的人。
今天先选最薄的一层。会议笔记 → 条目 → 可选日历草稿可以,但日历草稿必须人来确认。
为什么新手跳过
三个通道一起上像齐全;排障时分不清谁在动。先选一个通道跑通。
步骤 1——三层
- Skill = 剧本:这件事按什么步骤做、什么时候停。
- MCP = 门口:让 AI 按你批准的规则用外部工具或数据(邮件、文档、日历)。
- Plugin(插件)= 可安装的打包,里面可以有 skill,也可以带 MCP 连接:像一个应用包。
步骤 2——做饭比喻
用厨房比喻分清三层:别把「调料」和「炉灶权限」混成一件事。
| 层 | 厨房比喻 |
|---|---|
| Skill | 菜谱卡片 |
| MCP 工具 | 灶台、冰箱、外卖 App:真正的电器 |
| Plugin | 套餐盒:菜谱 + 电器优惠券 |
可以只用菜谱:把 skill 写在 Project 里就行。只有任务需要实时数据或真动作时,才打开电器(MCP),并且保持「行动前先问」。
步骤 3——新手顺序
- 1) 先把提示词套路写清:目标、格式、边界。
- 2) 存进 Project 说明或 skill 草稿。
- 3) 再连一个偏「只读」的安全 MCP 工具。
- 4) 需要现成维护包时再装插件,装前仍要审权限。
实例
填好的提示词
用一张新手能看懂的表解释:插件 vs skill vs MCP 连接器:各干什么、各举一个示例工作、主要风险是什么。
边界:不编造;未知标 [?];发送前停下。
完成检查:一分钟内能判过/不过。
像样的输出大概长这样
- 三行表:插件 / skill / MCP,各配一个日常例子
- 插件行点出「没读完的权限」这类打包风险
- skill 行停在剧本,不声称会自动发送
- MCP 行把写入/发送标成新手主坑
- 结尾写清「本周先碰哪一层」
- 一分钟内能对照清单判过/不过;缺项标红再继续
实例
填好的提示词
我想做:会议笔记 → 行动条目 → 可选日历草稿。请提一个最薄的 Skills/Plugins/MCP 栈,并写清暂时不接什么。
发送/删除默认关闭;先只读。
边界:不编造;未知标 [?];只出草稿。
像样的输出大概长这样
- 会议笔记 → 条目只用 skill / Project 说明
- 日历先停在对话里的草稿,不自动建事件
- 写明本月不接 Gmail 发送 / Drive 写入
- 栈短到一张便利贴,不画三年架构图
- 如果以后加连接器,仍强调行动前先问
保存这张卡
记下:今天选的最薄一层 + 本月绝不加的那一层。标上日期。
开始前
先说清本周必须打通的那件活。说不清活,就先不装。
本周怎么用
本周只选一个通道(plugin / skill / MCP)跑通一件真活;再记为何不选另外两个。
最薄层获胜
优先能完成每周工作的最小一层。真需要时再往上叠。
现实核对
把这篇指南放在真实任务旁边。说不出任务,就先挑一件再往下读。
一件活,一层就够
如果要用「并且」才能解释为什么装了某样东西,说明你叠早了。很多周里,只靠 skill 整理笔记就够;MCP 写日历是以后的升级,而且仍要「行动前先问」。
红旗
三个通道一起上,排障时分不清谁在动;先选一个通道跑通。
常见坑
| 坑 | 改法 |
|---|---|
| 三个概念一次全上 | 先选今天最薄的一层 |
| 日历草稿无闸 | 必须人确认再进真实日历 |
| 用长文解释替代对照表 | 一张表:干什么 / 例子 / 风险 |
现在就练一次
今晚就练:只选一个通道跑通一件真活;再记为何不选另外两个。
闭环
写下一行:下次三个通道一起上,我拒绝开跑。
一句话带走
先选一个通道跑通;再记为何不选另外两个。
下一步
← Back to home · Learn · Tools · Compare · Start