Claude Skill Engineering (Cross-Author Thematic Synthesis)

When to open: You have heard of Claude Skills and want to build one / want to engineer and package your own prompts / are choosing between Skill vs MCP vs Sub-Agent / want to know the three fatal pitfalls of Skills. This wiki fuses the perspectives of 6 author wikis on the Skill topic, and explicitly marks 4 core disagreements.

One-sentence core:Skill = a "prompt engineering product" that can be invoked by Claude / other Agents(Lao D's definition) = An Agent's wiki(Wang Kai's metaphor) = A collaborative carrier for Sequential Thinking + framework libraries + MCP(Lao Huang's practical experience).Skill is the key abstraction that upgrades prompts from "scattered text" to "software-grade assets".

0. Five perspectives on "Skill"

Perspective	Source	What is a "Skill"	One-sentence core
Product engineering school	newtype Skill system in practice (Super Analyst v1/v2 + Super Writer)(Huang Yihe)	Standard implementation of Anthropic Skill format(SKILL.md + references + scripts)	"An engineering paradigm for Sequential Thinking + MCP + framework libraries" (Super Analyst three-piece set)
Agent philosophy school	Wang Kai's Claude Code terminal workflow methodology (complete version) §4 (Wang Kai)	An Agent's wiki / internalizing work into reusable assets	"A Skill is an Agent's wiki / writing Skills in code minimizes information decay"
Psychotherapy school	Wang Kai's IFS Internal Family Systems AI-assisted practice + Claude Skill definition	Complete SOP + safety boundaries + trigger conditions	"IFS 7-step workflow + Schwartz 8C Self + three types of parts identification"
Creative application school	Xiaohongshu / AI tools + Skill tutorial collection	An SOP container that can directly call local services and splice files	"Skills not only turn work into SOPs, but can also call various local files and local services" (advantage over N8N)
Cross-platform reuse school	dontbesilent AI self-media course / 13-lesson systematic methodology §L13 (Lao D)	A cross-platform reusable prompt engineering product	"SKILL.md + references/ + Git version management / can be invoked by any Skill-compatible Agent"
Video production SOP school	Wang Kai's SillyTavern video production SOP (Skill implementation example)	A complete 7-stage engineering paradigm	"Phase 1-7: directory skeleton → asset fetching → Skill system → Components → first video → three types of videos → scaling"

1. Cross-author consensus (6 points)

1.1 Skill is "the end state of prompt engineering"

Lao Huang newtype · MCP Protocol and Ecosystem in Practice:"Skill-ification is the end state. Any prompt + workflow that is stable enough should be turned into a Skill".

Lao D corroborates (dontbesilent AI self-media course / 13-lesson systematic methodology §L13):

"Well-written system prompts are scattered across various platforms — LobeChat / Yuanqi / ChatGPT, with different effects across different models.Good prompts cannot be reused, and versions run wild. Skills solve this: centralized management of fragmented prompts (Git version control) + automatic trigger conditions + cross-platform reuse".

Wang Kai corroborates (Wang Kai's Claude Code terminal workflow methodology (complete version) §4.4):"If you want precision → write Skills in code". Text has high ambiguity; code/mathematics has the least information decay.

1.2 The three-piece structure of a Skill (Anthropic official + all-author consensus)

.skill = zip 包,含:
├── SKILL.md          (必有) Skill 核心定义 + workflow
├── references/       (可选) 详细文档(框架说明、用法示例、风格指南)
├── scripts/          (可选) Python / TypeScript 辅助脚本
├── README.md         (可选) 用户文档
└── INSTALLATION-GUIDE.md (可选)

SKILL.md frontmatter(All authors use the same format):

---
name: super-analyst
description: 简短描述 + 触发条件(LLM 自动识别何时调用此 Skill)
---

1.3 Boundaries between Skill vs MCP vs Sub-Agent

Dimension	Skill	MCP	Sub-Agent
Essence	Encapsulates Agent behavior patterns	Connects external tools	Launches a new Claude instance
Typical scenario	"How to think / how to search / how to make decisions"	"Call GitHub / databases / browsers"	"Scan different directories in parallel"
Invocation relationship	Can invoke MCP	Does not invoke Skills	Re-reads CLAUDE.md / Skills
Persistence	Permanent storage	Permanent storage	Single session

Lao Huang's judgment:Skill is the routing + selection + orchestration layer; specific content (prompts) is placed in MCP(can be independently maintained and reused).

1.4 Skill priority hierarchy (tested by Wang Kai)

Wang Kai's Claude Code terminal workflow methodology (complete version) §5.1:

Enterprise > Personal (User) > Project > Plugin
同名上层覆盖下层

Specific path:

Enterprise:/Library/Application Support/ClaudeCode/...(managed)
User (global):~/.claude/skills/<name>/SKILL.md
Project:./.claude/skills/<name>/SKILL.md
Local: Do not write a Skill (only write CLAUDE.local.md)

4-question decision(which layer to place it in):

Is this capability bound to a specific resource? (e.g., a certain browser port) → Project
Do different identities/agents have different credentials? → Project
Do different identities/agents behave differently? → Project
Pure logic, used identically by all agents? → User

1.5 Sequential Thinking is the "standard companion" for Skills (Lao Huang consensus)

newtype Skill system in practice (Super Analyst v1/v2 + Super Writer) §III Super Analyst Pro 2.0 uses Dual Sequential Thinking(dual scheduling):

Stage 2:Intelligence Planning(Sequential Thinking #1)
Stage 4:Framework Selection(Sequential Thinking #2)

newtype · MCP Protocol and Ecosystem in Practice §3.1:Sequential Thinking is the MCP that Lao Huang "installs and keeps by default".

Relationship with model-built-in Thinking:

The Thinking of flagship models (Claude 4 / o3 / Gemini 2.5 Pro) can replace it
but API access has costs → it helps models that do not have or do not enable Thinking
Best practice: equip Sequential Thinking and let the model decide for itself

1.6 Skill trigger conditions are automatically recognized by LLMs (Lao D + Lao Huang consensus)

dontbesilent AI self-media course / 13-lesson systematic methodology §L13:"SKILL.md = metadata + frontmatter + name + description + trigger conditions"

Example of Lao Huang's Super Analyst trigger conditions (newtype Skill system in practice (Super Analyst v1/v2 + Super Writer) §2.2):

description: Advanced analytical framework selector for complex problem-solving 
  and decision-making. Use when users need structured analysis, strategic thinking,
  problem diagnosis, or decision support. Automatically selects and applies the 
  most appropriate framework(s) from 12 specialized analytical methods...

Why trigger conditions are critical:

LLMs automatically judge "when to invoke this Skill" via the description
If trigger conditions are not written clearly → the Skill will never be invoked
If written too broadly → it will steal work from other Skills

2. Explicit disagreements (4 points — not smoothed over)

2.1 Disagreement 1: Should Skills be written in "natural language" or "code"?

Perspective	Position
Lao Huang Super Analyst	Primarily natural language + framework definitions — 539 lines of text in SKILL.md + 12 analysis frameworks (described in natural language)
Wang Kai(Wang Kai's Claude Code terminal workflow methodology (complete version) §4.4)	"If you want precision → write Skills in code" —— "Text has high ambiguity; code/mathematics has the least information decay. Scenarios requiring precision are very suitable for being converted into skills to ensure a unique effect"
Lao D(dontbesilent AI self-media course / 13-lesson systematic methodology §L13)	Hybrid — SKILL.md text + scripts/ code assistance

Wang Kai's practical demonstration:

"Various motion effects are often used during video production... If I tell Claude Code to follow instructions in text, it always takes various routes and cannot guarantee the output effect.Later I had it write code, put the knowledge points into code in the video skill, and the implementation path became relatively unique. "

Genuine disagreement:Trade-off between precision and flexibility:

Analysis frameworks / thinking processes → natural language (Lao Huang, allowing LLM flexibility)
Video motion effects / mathematical calculations / strict output formats → code (Wang Kai, ensuring uniqueness)
Content production SOPs → hybrid (Lao D)

2.2 Disagreement 2: Are Skills superior to orchestration frameworks such as N8N / LangChain?

Perspective	Position
Xiaohongshu vibecoding(Xiaohongshu / AI tools + Skill tutorial collection §3)	Skills are superior to N8N — "N8N is OK for making multiple requests to different models,but video splicing steps are cumbersome; Skills can directly call local splicing tools"
Lao D	Skill cross-platform advantage —— "Can be invoked by any Skill-compatible Agent"
Lao Huang(newtype miscellaneous practices: llms.txt protocol + CrewAI Obsidian project + n8n MCP workflow)	Skill + N8N are complementary — Expose N8N MCP Server Trigger as MCP, then invoked by Skills
Wang Kai multimodal Agent(Wang Kai's Claude Code terminal workflow methodology (complete version) §3.2)	Skill is the basic unit; engineering orchestration relies on multiple terminals + Codex CTO review

Xiaohongshu original quote:

"Skills can not only turn procedural, batch work into SOPs,but also very conveniently call various local files, and then use some very good local services. "

Genuine disagreement:Is Skill a substitute for N8N or an upper layer above it?:

Lao Huang's position: Skill is the upper layer; N8N is exposed to Skills via MCP for invocation
Xiaohongshu position: Skills directly replace N8N (because they can call local services)
Wang Kai's position: Skill is the basic unit; Codex CTO review + multiple terminals are the upper layer

2.3 Disagreement 3: Are more Skills always better?

Perspective	Position
Wang Kai(Wang Kai's Claude Code terminal workflow methodology (complete version) §4.3)	"More Skills lead to more conflicts" —— One of the fatal pitfalls: requires continuous structural organization → Karpathy LLM Wiki thinking(Organize all files, structure folders by project, hierarchy, and interconnections)
Lao Huang Super Analyst	Less but refined — 3 core Skills (Super Analyst v1 / Pro 2.0 / Super Writer) cover all analysis + writing scenarios
Lao D	Many but categorized — SKILL.md engineering prompt management, allows scattered storage, routed by trigger conditions

Wang Kai exclusive:"3 Fatal Pitfalls":

Rules must specify "scope of application" — Don't write "add tags later", write "When you post on Xiaohongshu, you must add tags" (platform-specific)
Memory cross-contamination — In one window talking about one thing, it replies with something completely unrelated from another project → check~/.claude/CLAUDE.md whether you wrote too many project names
More Skills mean more conflicts — Requires Karpathy LLM Wiki structured organization

2.4 Disagreement 4: At which layer should Skills be installed? (User global vs Project)

Perspective	Position
Wang Kai multi-matrix(Wang Kai Multi-Agent Browser Matrix Architecture (v1.1 Practical Guide) §4)	Most Skills installed at Project level (identity-level) — "publish-youtube" / "identity-A-tone" are all grouped by identity (rules specific to different YT accounts)
Lao Huang Super Analyst	Install at User level (global) — Usable across all projects
Lao D	No separation — Git version management, install on demand

Real disagreement:Matrix-based vs General-purpose:

Multi-identity operation / multi-account matrix → Wang Kai's position: Skills must be isolated at Project level
Single user / general tools → Lao Huang's position: shared globally at User level
Team / Git collaboration → Lao D's position: install on demand

3. Lao Huang Exclusive: Super Analyst Trio (Product-Grade Skill Template)

newtype Skill System in Practice (Super Analyst v1/v2 + Super Writer) Three complete Skill bundles:

3.1 Super Analyst v1(2025-Q3)

Core positioning:Intelligent selector for 12 analysis frameworks

12 frameworks:

First Principles Thinking
5 Whys
SWOT Analysis
Porter's Five Forces
Cost-Benefit Analysis
Design Thinking
Systems Thinking
Socratic Method
Pareto Analysis (80/20)
MECE Principle
Hypothesis-Driven Analysis
Scenario Planning

Key design:Framework prompts are fetched via prompt-house-local MCP(The Skill itself does not store the 12 complete prompts).

3.2 Super Analyst Pro 2.0(2025-10-26)

Slogan:Deep Thinking + Comprehensive Research + Structured Analysis = Consulting-Grade Insights

Upgrades (vs v1):

Dimension	v1	v2.0 (Pro)
Line count	180	539(3x)
Workflow	Simple selection	7-Stage Systematic Workflow
Thinking	None	Dual Sequential Thinking
Information	None	Web Search integration(Chinese-English bilingual)
Complexity	One-size-fits-all	Level 1/2/3 adaptive
Output	Framework results	Consulting-Grade Insights

7-Stage Workflow:

Stage 1: Problem Understanding & Assessment (10-30秒)
Stage 2: Intelligence Planning (Sequential Thinking #1) *[可跳]
Stage 3: Intelligence Gathering [可跳]
Stage 4: Framework Selection (Sequential Thinking #2) **[必跑]
Stage 5: Prompt Retrieval (via MCP)
Stage 6: Structured Analysis
Stage 7: Integrated Output

Level 1/2/3 complexity:

Level 1 Simple (1-2 minutes): concept explanation / framework description
Level 2 Medium (3-6 minutes): partial external information + 2-3 dimensions
Level 3 Complex (10-20 minutes): deep research + multi-dimensional strategy

3.3 Super Writer(2026-Q1)

3 major stacks:

Material Collection(Material collection)
7 Style Analysis types(Style analysis)
6 Writing Methodologies(Writing methodology)
Triple Sequential Thinking

4. Wang Kai IFS Skill Exclusive (Complete SOP + Safety Boundaries)

Wang Kai IFS Internal Family Systems AI-Assisted Practice + Claude Skill Definition —— Richard Schwartz IFS(Internal Family Systems) Psychotherapy framework made into a Skill:

4.1 IFS Core Concepts (Schwartz, 1980s)

Self — 8C qualities:

Calm / Curiosity / Compassion / Clarity / Confidence / Courage / Creativity / Connectedness

Three types of Parts:

Type	Function	Common manifestations
Managers	Proactive defense / avoiding hurt	Perfectionism / overplanning / self-criticism / control / people-pleasing
Exiles	Carry old wounds / repressed emotions	Inner child / deep fear / shame / abandonment / helplessness
Firefighters	Emergency firefighting	Binge eating / alcohol abuse / avoidance / angry outbursts / mindless phone scrolling / workaholism

4.2 IFS 7-Step Workflow (AI-Assisted Version)

准备:识别部分
  ↓
1. 找到它(Locate)
  ↓
2. 检查 Self 能量(Check for Self-Energy)— 最关键
  ↓
3. 靠近和建立联系(Approach & Connect)
  ↓
4. 倾听(Listen)
  ↓
5. 见证(Witness)
  ↓
6. 了解需求(Understand)
  ↓
7. 安置关闭(Unburden & Close)

4.3 Investment Scenario Application

Wang Kai applies IFS to:

Loss aversion / decision anxiety
Overconfidence / FOMO
Inner child (when facing losses)
Protector-Exile dynamics

This is an atypical Skill template —— Psychological work SOPs can also be turned into Skills, as long as there is a structured process + safety boundaries.

5. Wang Kai SillyTavern Video Skill Implementation Example (Phase 1-7)

Wang Kai SillyTavern Video Production SOP (Skill Implementation Example) —— Complete 7-stage engineering template, transferable to mass production of any product video:

Phase	Content
1. Directory skeleton	Video project structure / assets / Components / Compositions layering
2. 8 dimensions of reference video analysis	Topic / shots / pacing / emotion / theme / style / BGM / transitions
3. 5 categories of real asset capture	Screenshots / characters / scenes / objects / Voice
4. Skill system / Components / Compositions layering	Reusable units → composition
5. First video	MVP validation
6. Three types of videos + Jimeng mixed editing	Content matrixization
7. Scaling	Replication + Skill optimization

6. Lao D L13 Skill Engineering Management (Exclusive Domestic Adaptation)

dontbesilent AI self-media course / 13-lesson systematic methodology §L13:

6.1 Pain points

Well-written system prompts are scattered across various platforms:

One version on LobeChat
One version on Yuanqi
One version on ChatGPT
Different models produce different effects

Result: Good prompts cannot be reused, versions run wild.

6.2 Claude Skill solution

Skill = a "prompt engineering product" that can be invoked by Claude Code / other Agents:

SKILL.md = metadata + frontmatter + name + description + trigger conditions
references/*.md = Long documents are not written in the main SKILL file, but are automatically loaded by Claude when the Skill triggers
scripts/ = Supporting tools (scripts / data files)

Advantages:

Centralized management of fragmented prompts(Git can be used for version control)
Automatic trigger conditions(similar to function signatures)
Cross-platform reuse(Any Skill-compatible Agent can invoke it)

6.3 Practical workflow

Organize the prompts you have made into SKILL.md
Use Claude 4.5 to organize frontmatter (name / description trigger conditions)
Split long references into references/*.md
Test trigger conditions
Git commit = version management + team sharing

6.5 zara Zhang Zala Skill Data Distribution Model: Centrally fetch, locally Remix (2026-05-18 follow-builders practice)

zara Zhang Zala / follow-builders skill and AI information source curation methodology Added a new dimension not covered in L13 new dimension:When a Skill's workflow depends on external data (API calls / remote content), should users get their own keys or should the author provide them centrally?

(Author: zara Zhang Zala, GitHub: zarazhangrui. The skill is open source github.com/zarazhangrui/follow-builders)

6.5.1 Pain points (v1 user self-crawling failure)

zara's early follow-builders v1: Each user's Agent independently called the Supadata API (YouTube transcripts) + X official API (Twitter) → users gave up at the first step because:

X API is paid and configuration is complex
Most users are unwilling/unable to get an API key

6.5.2 Solution: Centrally fetch, locally Remix

作者侧(中心化):
  每日脚本 → 调 API(作者付费) → 抓内容 → 写入 JSON 数据源

用户侧(skill 装在 OpenClaw / Agent):
  skill 触发 → fetch 中心 JSON → 按用户偏好本地 remix(语言/长度/风格) → 推送

6.5.3 Model comparison

Dimension	v1 user self-crawling	v2 central fetch + local remix
User barrier	High (get own API key + pay)	Zero (install skill and use immediately)
API cost	Borne by users	Borne by author (controllable scale)
Content quality	Determined by user-written prompts	Curated by author, quality guaranteed
Personalization	Both content and format are personalized	Unified content, format remixed locally
Value proposition	Tool	Curation + tool

6.5.4 Mindset spillover

Any product form of "author-maintained external data + client-side skill invocation" can apply this: industry briefings / professional knowledge bases / dataset distribution (stocks/house prices/recruitment) / paid API relay (can be combined with subscription monetization).

Core quote(zara): "The greatest value of a skill is the high-quality information sources after careful curation" — The value lies not in the distribution mechanism, but in the curation itself.

6.6 zara Zhang Zala Skill Product Portfolio Methodology (2026-05-22 full sweep)

See zara Zhang Zala / Skill & Independent Product Portfolio (Strategy Review). Expanded from the single point of follow-builders to zara's full year of Skill practice system.

6.6.1 Numerical results (product-grade Skill ranking evidence)

Skill / Product	Data	Ranking
Frontend Slides	GitHub 16,000+ star(2026-05)	#1 Skill in the PPT domain
follow-builders	Open source + built-in Skill repository	See §6.5
Feishu CLI agent skill	Open sourced	Meeting todo → done automation
Long-to-short video skill	Open sourced	Automatically extract highlight clips

6.6.2 Key differences from the first 5 perspectives (Lao Huang / Wang Kai / IFS / xhs / Lao D)

Dimension	First 5 perspectives	zara 7th perspective (opinionated distribution school)
Skill value positioning	Engineering encapsulation / SOP reuse / safety boundaries	Carrier of the author's worldview / opinions
User-Skill relationship	Tool invocation	Accept the author's brainwashing (observed as a feature)
Plagiarism risk	"Your prompts being stolen"	Cannot be plagiarized — what is copied is the tool, not the opinions
Business model	Paid / embedded ads	Free, product = self-expression, traffic brings 260k Xiaohongshu followers

6.6.3 Skill > SaaS (zara's contrarian choice)

Most of zara's products Are first made as Claude Skills (open source on GitHub) rather than web apps:

Option	Skill form	SaaS form
Deployment cost	0(GitHub README)	Low (README + one post to 260k followers)
Distribution cost	0 High (SEO / PH / advertising)	User modification
User modification	fork(Everyone has their own version)	Not modifiable
Maintenance	Writing documentation is enough	24h online + bug fixes
Coding ≠ Shipping; 70% of Shipping work	Not needed	Must be done(Excalicord personal case: 8 systems integrated)

zara mindset(See zara Zhang Zala / Skill & Independent Product Portfolio (Strategy Review) §4):"Building products for yourself vs building products for others, the difficulty is not on the same order of magnitude. Skill is an intermediate form between 'pure local tool' and 'end-to-end SaaS' — enjoying the lightness of a tool + users being able to use your work."

6.6.4 Fork > PR (zara exclusive)

zara does not want users to submit PRs (to modify zara's main repo) for open source skills such as follow-builders / Tab Out

"In this day and age, everyone can modify their own version. I want you to fork fork my product, not submit a PR."

Mechanism: Skill is opinionated → everyone has different aesthetics → it is suitable for everyone to fork a copy and modify their own version > everyone uses the same product.

→ This point was not explicitly stated in any of the previous 5 perspectives.

6.6.5 Spillover of Mental Models (The Maximum Leverage for Builder vs Influencer)

zara ties Skill engineering together with content creation:

做 Skill(Builder)
   ↓
26 万粉小红书发示范作品(Influencer)
   ↓
GitHub star 涨(数据飞轮)
   ↓
更多用户 fork(自我筛选 + 信任沉淀)
   ↓
更多 Skill idea(Twitter 看技术 + Claude Code brainstorm)
   ↓
循环

For details see zara Zhang Zala / Content Creation SOP (validated by 260k Xiaohongshu followers + 50k Twitter followers) §5.4 build in public + zara Zhang Zala / PM & VC career trajectory + Computer History Museum docent §7 The intersection of Builder + Influencer = maximum leverage.

7. Xiaohongshu Vibecoding Skill in Practice (Exclusive Creative Applications)

Xiaohongshu / AI Tools + Skill Tutorial Collection:

7.1 IELTS 7.5 Skill Three-Piece Set (No Paid Classes)

IELTS Diagnose Skill(Diagnosing speaking / writing / reading)
IELTS Speaking Skill + Gemini Live
IELTS Writing/Reading Skill + Notebook LM

Complete SOP for self-study to 7.5.

7.2 Vibe Marketing Skill (Batch Video Production)

Why Skill is better than N8N:

N8N	Skill
Making multiple requests to different models (image/text/video) is OK	Same capability
Video splicing steps are cumbersome	Can directly call local splicing tools
Workflow visualization	SOP-ification + local resource invocation

7.3 Obsidian Skill Diagram Drawing

Skill made by the founder of Obsidian:Draw H16Z long-form logic diagrams in 10 seconds.

7.4 Pixel Sprite Agent (VS Code Extension Visualization)

"Cyber boss supervising AI at work experience" — not a Skill, but demonstrates the possibility of Skill / Agent visualization.

8. How to Choose — Skill Paths by Identity

Who you are	Primary method	Starting point
Want to build product-grade Skills	Lao Huang Super Analyst route	7-Stage Workflow + Dual Sequential Thinking + fetching prompts via MCP
Want to build psychology / thought process Skills	Wang Kai IFS Skill route	Complete 7-step SOP + safety boundaries + trigger conditions
Want to batch produce videos / content	Wang Kai SillyTavern 7-stage / Xiaohongshu Vibe Marketing	Components + Compositions layering + local service invocation
Want to manage scattered prompts across platforms	Lao D L13	SKILL.md + references/ + Git version control
Want precision scenarios	Wang Kai "write Skills with code"	Write Python / TypeScript code in scripts/
Multi-identity operation	Wang Kai Project-level Skill	Independent.claude/skills/ for different identities

General starting sequence(for any identity):

Write SKILL.md frontmatter(name / description / trigger conditions)
If content is long → split into references/*.md(automatically loaded when Skill is triggered)
If precision is needed → write code in scripts/(minimal information decay)
Test trigger conditions(whether the LLM can automatically recognize "when to invoke")
If calling external tools → call MCP instead of hardcoding in the Skill
Add Sequential Thinking scheduling(essential for complex tasks)
Git commit + version control
Pass the 3 fatal pitfall checks(Wang Kai): Do the rules have a "scope of application"? Will memory cross-contaminate? Will there be conflicts when there are many Skills?

9. Reference List

Main original author wikis:

newtype Skill System in Practice (Super Analyst v1/v2 + Super Writer) —— Super Analyst v1/Pro 2.0/Super Writer / 12 frameworks / 7-Stage / Dual Sequential Thinking / Anthropic official format
Wang Kai Claude Code Terminal Workflow Methodology (Complete Edition) §4 —— 3 fatal pitfalls / writing Skills with code / Skill priority hierarchy / 4-question decision making
Wang Kai IFS Internal Family Systems AI-Assisted Practice + Claude Skill Definition —— Schwartz 8C Self / three types of parts / 7-step workflow / investment scenario application
Wang Kai SillyTavern Video Production SOP (Skill Implementation Example) —— Complete engineering of Phase 1-7 / Components + Compositions layering
Xiaohongshu / AI Tools + Skill Tutorial Collection —— IELTS three-piece set / Vibe Marketing Skill / pixel sprite / Obsidian diagram-drawing Skill
dontbesilent AI Self-Media Course / 13-lesson Systematic Methodology §L13 —— SKILL.md + references/ + Git / domestic adaptation / cross-platform reuse
zara Zhang Zala / follow-builders Skill and Curated AI Information Source Methodology —— Centrally fetch, locally Remix data distribution mental model(follow-builders practice 2026-05-18)/ curation + centralized JSON + local remix

2026-05-19 Supplement: Anthropic pushes Agent Skills as an open standard

From equipping-agents-for-the-real-world-with-agent-skills (first released 2025-Q4, published 2025-12-18 as open standard agentskills.io). This is the official schema implementation, cross-checked against the "Skill practice" of the previous 6 author wikis.

What is a Skill (Anthropic official definition)

"organized folders of instructions, scripts, and resources that agents can discover and load dynamically to perform better at specific tasks"

Translation: An organized folder containing instructions, scripts, and resources that an agent can dynamically discover and load to perform better on specific tasks.

Anthropic's own life analogy:Writing a skill for an agent is like writing onboarding documents for a new employee.

Official schema(consistent with §1 of this wiki):

my-skill/
├── SKILL.md           # 主指令 (YAML frontmatter + description + 步骤)
├── references/        # 参考资料 / 范例 / 数据
└── scripts/           # 可执行脚本 (Python / Shell)

Why is Skill needed as a new primitive? (Official explanation)

The previous 3 ways to add domain expertise to agents all had flaws:

Method	Drawbacks (listed by Anthropic)
Stuffing into system prompt	The more you stuff, the longer it gets → pollutes default behavior + context explosion
Custom fine-tuning	Expensive / slow / not portable
MCP server	Heavy + suitable for "calling external systems", not "installing domain knowledge"

Positioning of Skill:Lightweight, dynamically loaded, portable —— does not pollute the agent's default behavior at ordinary times, only loaded in trigger scenarios.

Boundary with MCP (officially clarified)

	Skill	MCP
What is provided to the agent is	Documentation + scripts + examples (domain knowledge)	Tool interfaces (calling external systems)
Trigger	The agent decides to invoke on its own	The agent sees it in the tool description → selects it
Loading	Dynamic import + unload	Connected at startup, stays attached long-term
Cross-platform	Yes (open standard)	Yes (open standard)
Content	Mainly SKILL.md natural language + scripts	APIs exposed by server code

Anthropic explicitly states:the two are complementary and not substitutes; in actual projects both are used — Skill holds "domain workflows", MCP holds "external system access".

Key cross-check with the previous 6 author wikis

Topic	Three practitioner camps	Anthropic official
Does Skill use natural language or code?	Wang Kai "write Skills with code" (advocates code) / Lao Huang Super Analyst takes natural language prompt route	The official schema is SKILL.md (natural language) + scripts/ (code),the two coexist
Skill vs MCP?	Lao Huang "Skill is getting stronger" / Wang Kai "Skill priority"	Official:complementary and not substitutes, Skill is domain knowledge, MCP is system access
Dynamic discovery?	xhs three-piece set hands-on practice / dontbesilent SKILL.md cross-platform	Official core feature: discover and load dynamically —— the agent itself finds the skill to use based on the task
Open standard?	(no corresponding entry)	2025-12-18 published agentskills.io, cross-platform portable

Other Skill-related clues from Anthropic Engineering's 25 articles

Anthropic Engineering Team First-Party Perspective: Agent Engineering / Skills / MCP / Claude Code (2025-2026 Complete Collection) §5 complete Skill chapter
harness-design-long-running-apps uses frontend-design SKILL as one part of the long-running harness
The 5 principles of "tool description prompt-engineering" given in writing-tools-for-agents can be directly applied to writing SKILL.md

History

2026-05-22:Supplement §6.6 zara Skill product portfolio methodology (based on a full sweep of 474 posts from zara's corpus). Expanded from the single point of follow-builders to zara's 7th perspective "opinionated distribution school": Skill is a carrier of opinions, not a tool / Skill > SaaS selection / Fork > PR / numerical evidence such as 16k+ star Frontend Slides / Builder + Influencer intersection = maximum leverage. Source wiki: zara Zhang Zala / Skill & Independent Product Portfolio (Strategy Retrospective) / Zara Zhang Zala / Content Creation SOP (Validated on 260k-follower Xiaohongshu + 50k-follower Twitter) / Zara Zhang Zala / Vibe Coding Mental Model System
2026-05-19: Added §2026-05-19 section (Anthropic promotes Agent Skills as an open standard). Added official definition, official schema, clarified boundary with MCP, cross-checked against the previous 6 author wikis. Added cross-link to Anthropic Engineering Team First-Party Perspective: Agent Engineering / Skills / MCP / Claude Code (2025–2026 Complete Collection).
2026-05-18: Topic-dimension wiki first release. Synthesized and rewritten from 6 author wikis. Explicitly marked 4 points of divergence: Skill natural language vs code / Skill vs N8N / Are more Skills better? / Which layer to install: User vs Project
2026-05-18 (evening): Added §6.5 Skill data distribution model "Centrally fetch, locally Remix" (source: Zara Zhang Zala follow-builders skill practical sharing, 2026-05-18 video / GitHub: zarazhangrui). Covers a new dimension not discussed in L13: for skills that depend on external data, should users provide keys or should the author centralize.(Initial version incorrectly attributed the author to Lao D; corrected after user feedback on 2026-05-18)