Chapter 5 — Load a Skill You'll Actually Use

What Is a Skill

WorkBuddy itself handles understanding tasks and organizing execution; a Skill is a reusable set of instructions, scripts, reference material and resources that tells the Agent how a certain kind of task should be done, which tools to call, and what format to deliver.

Anthropic officially launched Agent Skills in October 2025, and published them as an open standard in December 2025.

A most standard Skill looks roughly like this:

Plain

my-skill/
├── SKILL.md
├── scripts/
│   └── check.py
├── references/
│   └── guide.md
└── assets/
    └── template.pptx

Only SKILL.md is required.

Markdown

---
name: tech-article-writing
description: For writing articles about AI products, model reviews and the tech industry
---

When you receive a writing task:

1. First confirm the article's core angle
2. Find primary sources
3. Cross-check core facts
4. Complete the draft in the user's writing style
5. Check banned phrasing and AI-sounding expressions

You can also include:

Plain

references/style.md

How a Skill Works

The most important design behind a Skill isn't actually SKILL.md — it's Progressive Disclosure.

Say your Agent has 100 Skills installed.

It doesn't dump all 100 Skills' full content into the context at once. That would waste tokens and flood the model with irrelevant instructions.

The standard approach has three layers.

First layer: at Agent startup it only looks at every Skill's name and description.

For example:

Plain

pptx
Handles PowerPoint create, edit, read tasks

pdf
Handles PDF extract, merge, edit, fill tasks

tech-article-writing
Writes AI and tech-industry articles

Second layer: when the user says:

Plain

Write a WeChat article about WorkBuddy

The Agent judges from the description that tech-article-writing may be relevant, and only then loads the full SKILL.md.

Third layer: during execution, if it finds it needs to imitate your writing style, it then reads:

Plain

references/style.md

If it needs to check for AI-sounding text, it then runs:

Plain

scripts/check-ai-phrases.py

The standard recommends that all Skills load only tens to hundreds of tokens of metadata at startup, load full instructions after a Skill is activated, and keep loading other material and scripts on demand. OpenAI Codex uses a similar mechanism: it first exposes Skill names, descriptions and paths to the model, then reads the full content when the model decides to use it.

So a Skill solves a long-standing Agent problem:

How to give an Agent lots of knowledge and working methods without stuffing everything into the prompt forever.

What's the Difference Between a Skill and a Prompt

This is the core question.

Dimension	Prompt	Skill
Core role	Describes the current task	Defines how a class of tasks is done
Lifecycle	Usually for one request	Long-term reusable
Trigger	User actively enters it	Agent auto-selects or user explicitly invokes
Carrier	Mostly text	A folder
Content	Instructions, context, examples	Instructions, scripts, material, templates, resources
Context cost	Usually enters context directly	Loaded on demand
Reuse	Often copy-pasted	Natively reusable
Sharing	A prompt text	A complete capability pack
Execution	Itself only instructions	Can call bundled scripts and tools
Model params	Doesn't change	Doesn't change either

The simplest way to understand it:

Plain

Prompt = the task
Skill = how to do it

What Skills Are Good For

The first use is giving the model procedural knowledge. Large models know a lot, but they may not know how your specific thing should be done — for example it knows SQL, but not your company's:

Plain

Which table holds the canonical user_id
The subscriptions table is append-only
Refund queries must exclude a certain status
What the corresponding Grafana dashboard ID is

This kind of knowledge is perfect for a Skill. Anthropic uses hundreds of Skills internally and found they cluster mainly in nine scenarios: API and internal library usage, product validation, data analysis, business-process automation, code scaffolding, code review, CI/CD, incident runbooks and infrastructure ops.

The second use is locking down complex workflows, like running an industry research pass.

A plain prompt might be:

Plain

Do a detailed research pass on WorkBuddy

The model re-thinks every time:

Plain

Where to find materials
What to check first
How to validate
Who to compare against
What structure to output

A Skill can fix the flow:

Plain

1. Official website
2. Official WeChat account and launch events
3. Product docs
4. Actual product testing
5. Comparison with similar products
6. Core-point distillation
7. Fact-checking

This kind of capability is called an Encoded Preference Skill. The model could already do each step individually, but the Skill organizes those steps the way a team or individual works.

Another kind is a Capability Uplift Skill — it supplements a capability the model can't do well or stably, such as complex document, PDF and PPT handling.

The third use is cutting down repeated prompts.

When you work with AI now, a lot of what you say is repeated. For example you often tell me:

Plain

Don't write too AI-sounding
Mix long and short sentences
Don't over-use bullet lists
Have your own judgment
Be restrained with technical content
Don't make up examples

These are naturally suited to a writing-style Skill.

From now on your prompt only needs:

Plain

Write a WorkBuddy article

Writing habits, source standards, banned expressions, article flow — all provided by the Skill.

The fourth use is turning personal and organizational experience into assets.

The biggest problem with traditional prompts is they tend to scatter across:

Plain

Chat history
Feishu docs
Notion
Someone's head

A Skill is a file, so it can:

Plain

Be Git-managed
Roll back versions
Be shared across a team
A/B tested
Auto-evaluated
Continuously updated

This matters a lot.

Find a Suitable Skill in WorkBuddy

Open "Experts · Skills · Connectors" on the left; you can search the skill market or describe your need with "Find a skill".

You can also find a suitable Skill in the SkillHub market.

Besides installing directly from the recommended list, you can also import a skill you downloaded yourself.

For example, you find a useful skill pack online and download a zip; the flow is: click "Upload Skill" and load the zip.

Use a Skill to Solve a Task

For example, you had the AI write an article and need to remove the AI-sounding tone. Find the "Article de-AI tool" Skill, install it, and when using it, type "/" to bring it up.