1.5 Skill: Add Your First Reusable Capability
A familiar situation: every time you ask the agent to write a product description, you restate the same requirements — brand voice, fabric composition, care instructions, SEO structure, three languages.
Once or twice is manageable. Ten times a week, five hundred times a year, every repetition is wasted work, and because the wording drifts each time, so does the output.
A Skill is built for exactly this kind of repetition.
What a Skill Is
A Skill is a reusable set of operating instructions: it defines which tools to use for a class of task, in what order, and in what format to deliver.
Its division of labour with a prompt fits in one line: a prompt covers "this time"; a Skill covers "every time".
| Dimension | Prompt | Skill |
|---|---|---|
| Core purpose | Describe this one task | Define how a class of task is done |
| Lifespan | Usually one-off | Reusable over the long term |
| Trigger | You type it | The agent selects it as needed, or you invoke it explicitly |
| Form | Text | A folder: instructions, scripts, references, templates |
| Context cost | Enters context directly | Loaded on demand, no cost while inactive |
| Reuse | Rewritten each time | Reusable by design |
| Blast radius of one edit | This task only | Every subsequent run |
| Can carry scripts | No | Can bundle scripts and templates |
The last two rows are the differences that matter in practice. Blast radius decides where knowledge accumulates: correct a prompt and you correct it again next time; correct a Skill and you never repeat it. Script support decides the ceiling: a Skill can ship a validation script and an output template, so it can do more than plain instructions.
On disk, a Skill is a folder, and the minimum viable version is a single SKILL.md:
---
name: product-description
description: Write multilingual product descriptions for apparel SKUs.
---
1. Read the SKU's fabric, composition, and care label.
2. Apply the brand voice guide.
3. Structure the title and bullets for SEO.
4. Output English, French, and German versions.
5. Check for banned phrasing before delivery.The structure is covered in 2.2. This lesson is about getting one installed and running.
Adding One: Three Steps
For the complete flow of adding a Skill, selecting it in a task, and reviewing the result, see AOS Work: Skill.
Explore is the single entry point for every addable resource. A Skill has the shortest flow, and it serves as the template for the other three types.
- Choose. Open 【Explore】 in the left sidebar and switch to 【Skills】. Each card carries a description, and opening it shows what is included.
- Add. Click 【Add】, check the name and included content, then confirm. Most Skills install directly; a few need extra authorization or configuration, and the page will say so.
- Confirm the state. Go back to the card and check that it reads 【Enabled】.
Use 【Includes】 on the detail page to review the packaged skills. The 【Enabled】 switch controls the resource's enabled state.
Invoking It in a Task
An enabled user-level Skill enters your context by default — you do not select it each time. When the task matches its description, the agent loads it automatically. Skills packaged inside a Work are plugin-level and do not work this way; see 1.6.
To be explicit, click 【Add context】 in the input box (or type /), select the Skill, confirm that a context tag appears in the input area, then send the task.
View Calls and Output
Use these locations to see the Skill used in the task and its output:
| Check | What you will see |
|---|---|
Type / to open the selection list | The Skill appears in the list |
| Select it explicitly and send | The Skill tag appears in both the input and the sent message |
| Run a scoped practice task | The result follows the Skill's steps and output format |
If the Skill is selected but the result misses expectations, check the inputs, steps, and output requirements in turn. Lesson 2.4 covers adjustments.
When a Skill Is Worth Installing
A Skill's value comes from reuse. One test: will this kind of task come back in a similar form?
| Situation | Choice |
|---|---|
| Weekly work with a stable method | Install an existing Skill, or build one |
| A one-off that will not recur | Write the requirements into the prompt |
| The method is still settling | Run it as a prompt a few times, then package it |
Summary
A prompt covers "this time"; a Skill covers "every time".
You have now walked the general flow once: add in Explore → confirm enabled → invoke in a task → check the output. The next lesson, Work, repeats that flow with one extra step: after installing, you also select it in the task.
1.4 Resource Map: The Four Kinds of Resources in Explore
The rest of the book uses these names directly. This lesson sets out their types, their sources, and how each one takes effect.
1.6 Work: Combine Several Skills Into One Work Package
A single Skill covers one class of task. A full business chain needs several, and Work fixes that combination in place.