Skip to main content
Updated Feb 10, 2026

Teaching Your Employee to Write

In the previous lessons, you connected to your AI Employee through Telegram and gave it tasks. The AI responded based on its general training. Now you're going to teach it something specific to YOU. Something it will remember every session. Something you can take with you to any platform.

This is the moment your AI Employee stops being a generic assistant and becomes YOUR specialized worker.

When you hire a human employee, you don't just assign tasks. You train them. You teach them how YOUR company writes emails, how YOUR team formats reports, what tone to use with YOUR clients. That training is what makes them valuable. The same principle applies to AI Employees, but with one crucial difference: the training you create is portable. Build it once for OpenClaw, and it works in Claude Code, Goose, or any AgentSkills-compatible platform.


What Are Skills?

Skills are expertise packages. They're markdown files with a specific format that tell your AI Employee how to handle particular types of work. When you create a skill for "writing professional emails," you're encoding exactly how you want emails written: the tone, the structure, the questions to ask before drafting.

Think of skills like training manuals for a new hire:

  • Without training: Employee uses generic approaches, requires constant correction
  • With training: Employee follows your standards, produces consistent quality

The technical format is simple: a directory containing a SKILL.md file with YAML frontmatter at the top (metadata) and instructions below (the actual training).


Where Skills Live (Precedence)

OpenClaw loads skills from three locations, and if names conflict, it uses the first one found:

PriorityLocationUse Case
Highest<workspace>/skills/Your current project's specific skills
Middle~/.openclaw/skills/Your personal skills (available to all agents)
LowestBundledBuilt-in skills shipped with OpenClaw

This precedence system means:

  • Workspace skills override everything (project-specific customizations)
  • Personal skills apply everywhere you work (your standard toolkit)
  • Bundled skills provide defaults you can override

For your first skill, we'll use the personal location (~/.openclaw/skills/) so it's available in every session.


Hands-On: Create Your Email-Drafter Skill

Time to build something real. You'll create a skill that makes your AI Employee an expert at drafting professional emails.

Step 1: Create the Skill Directory

Open your terminal and run:

mkdir -p ~/.openclaw/skills/email-drafter

Output:

(no output means success)

This creates the directory structure OpenClaw expects. The directory name (email-drafter) will be how you reference this skill.

Step 2: Create the SKILL.md File

Create the file ~/.openclaw/skills/email-drafter/SKILL.md with this content:

---
name: email-drafter
description: Draft professional emails with appropriate tone and formatting. This skill should be used when users need to write emails, compose messages, or communicate professionally via email.
metadata: { "openclaw": { "always": true } }
---

# Email Drafter

You are an expert email writer. When asked to draft an email, follow this process:

## Information Gathering

Before drafting, confirm you have:

1. **Recipient**: Who is this email to? (colleague, client, manager, external)
2. **Purpose**: What's the goal? (request, follow-up, introduction, thanks, apology)
3. **Key Points**: What must be included?
4. **Tone**: Formal, professional, friendly, or urgent?

If any information is missing, ask for it before proceeding.

## Email Structure

Always include these elements:

- **Subject line**: Action-oriented, clear, under 60 characters
- **Greeting**: Appropriate to relationship and culture
- **Opening**: State purpose in first sentence
- **Body**: Supporting details, context, specifics
- **Call to action**: What do you want the recipient to do?
- **Closing**: Professional sign-off appropriate to tone

## Tone Guidelines

| Tone | Greeting | Language | Closing |
|------|----------|----------|---------|
| **Formal** | "Dear Mr./Ms. [Name]" | "I am writing to..." | "Best regards" |
| **Professional** | "Hi [First Name]" | Direct, clear | "Thanks" or "Best" |
| **Friendly** | "Hey [First Name]" | Conversational | "Cheers" or "Talk soon" |
| **Urgent** | "Hi [Name]" | "Time-sensitive:" prefix | "Please respond by [date]" |

## Output Format

Present every drafted email in this format:

Subject: [Clear, action-oriented subject]

[Greeting],

[Opening - state purpose immediately]

[Body - supporting details]

[Call to action - be specific about what you need]

[Closing], [Sender name placeholder]


After the first draft, always offer: "Want me to adjust the tone, length, or emphasis?"

Step 3: Verify the Skill Format

Let's confirm the structure is correct:

cat ~/.openclaw/skills/email-drafter/SKILL.md | head -20

Output:

---
name: email-drafter
description: Draft professional emails with appropriate tone and formatting. This skill should be used when users need to write emails, compose messages, or communicate professionally via email.
metadata: { "openclaw": { "always": true } }
---

# Email Drafter

You are an expert email writer. When asked to draft an email, follow this process:

## Information Gathering

Before drafting, confirm you have:

1. **Recipient**: Who is this email to? (colleague, client, manager, external)
2. **Purpose**: What's the goal? (request, follow-up, introduction, thanks, apology)
3. **Key Points**: What must be included?

Testing Your Skill

Restart to Load the Skill

Skills are loaded when a session starts. To pick up your new skill, you need to start a fresh session.

If using the OpenClaw gateway:

# Restart the gateway to reload skills
openclaw gateway restart

Or simply start a new conversation with your AI Employee via Telegram.

Test the Skill

Send this message to your AI Employee:

Draft an email to my manager requesting a meeting to discuss my career development goals.

What you should see: Your AI Employee now asks clarifying questions based on your skill's "Information Gathering" section before drafting. It follows the structure you defined. It offers to adjust after the first draft.

Compare this to how it would have responded before: a generic email with no systematic approach.

Verify Skill Loading

You can check which skills are loaded:

openclaw skills list

Output:

Loaded skills:
- email-drafter (personal: ~/.openclaw/skills/)
- [other skills...]

If email-drafter doesn't appear, check:

  1. Directory name matches name: in frontmatter
  2. File is named exactly SKILL.md (case-sensitive)
  3. YAML frontmatter is valid (no syntax errors)

Understanding the Skill Format

Let's break down what makes a skill work:

YAML Frontmatter (Required)

---
name: email-drafter
description: Draft professional emails with appropriate tone...
metadata: { "openclaw": { "always": true } }
---
FieldPurposeNotes
nameUnique identifierLowercase, hyphens only, must match directory name
descriptionWhen to use this skillHelps AI decide if skill applies; include trigger phrases
metadataPlatform-specific settings"always": true means skill loads every session

Instruction Body

Everything after the frontmatter is the actual training. Write it like you're explaining to a capable new employee:

  • What to do first (information gathering)
  • How to structure the output
  • Guidelines for variations (tone, format)
  • What to offer after completion

The Portability Promise

Here's where skills become powerful: they're not locked to OpenClaw.

The AgentSkills Standard

OpenClaw uses the AgentSkills format, which is an open standard. Any platform that supports AgentSkills can use the same skill files. This includes:

  • OpenClaw (what you're using now)
  • Claude Code (Anthropic's CLI)
  • Other AgentSkills-compatible tools

Test Portability: Claude Code

If you have Claude Code installed, test the same skill there:

# Copy to Claude Code's skill location
mkdir -p ~/.claude/skills/email-drafter
cp ~/.openclaw/skills/email-drafter/SKILL.md ~/.claude/skills/email-drafter/

Now in Claude Code, the same skill works identically. One skill, multiple platforms.

Why Portability Matters

Without PortabilityWith Portability
Train AI in each platform separatelyBuild once, use everywhere
Different behavior across toolsConsistent quality everywhere
Lost work when switching platformsYour skills are YOUR assets
Vendor lock-inPlatform independence

The skills you build are intellectual property. They encode your expertise, your standards, your way of working. They're portable because they should belong to YOU, not to any platform.


What You Built

In this lesson, you:

  1. Created your first skill - A working email-drafter that follows YOUR standards
  2. Understood skill structure - YAML frontmatter + instruction body
  3. Learned skill precedence - Workspace > Personal > Bundled
  4. Verified portability - Same skill works across platforms

This is the foundation of the Digital FTE paradigm. You're not just using AI. You're training AI with reusable, portable expertise that compounds over time.


Try With AI

Prompt 1: Test Your Email Skill

Draft a professional email to a potential client introducing our consulting services. The client is a mid-size tech company. Keep it under 150 words.

What you're learning: You're testing whether your skill's structure (information gathering, format, tone guidelines) actually influences the output. Compare the response to what you'd get from a generic AI without the skill.

Prompt 2: Test Edge Cases

I need to follow up on an unanswered email from last week. The person is a senior executive. Draft something polite but firm that creates urgency without being pushy.

What you're learning: Edge cases reveal whether your skill handles nuance. The tension between "firm" and "not pushy" requires the AI to apply your tone guidelines thoughtfully. If the result isn't right, you've found an area to improve in your skill.

Prompt 3: Test Adaptation

Draft an email declining a meeting invitation politely while suggesting an alternative time next week.

What you're learning: Declining requires a different structure than requesting. You're testing whether your skill is flexible enough to handle the full range of email types, or whether you need to expand its guidelines.

Safety reminder: When testing email skills with real scenarios, avoid including actual confidential information. Use realistic but fictional details, or anonymize sensitive content before sharing with your AI Employee.