Skip to main content

Teach Your Employee a Skill

In the previous lesson, you gave your employee an identity. It knows who it is, what rules to follow, and where its boundaries are. But send it a real professional task and you will notice something: it responds with general knowledge, not expert judgment.

Ask an accountant-employee to review an invoice and it gives you a generic checklist. Ask a teacher-employee to plan a lesson and it produces something any search engine could generate. The identity tells the employee WHO it is. A skill tells it HOW to think about specific tasks -- the decision rules, the things to check, the edge cases that separate a professional from an amateur.

This lesson's challenge: encode your most valuable professional expertise into a skill that makes your employee think like a domain expert, not a well-read generalist.

Skills Are Manufactured Assets

When you write a SKILL.md file, you are not configuring a chatbot. You are manufacturing a portable capability: version-controlled, shareable, and deployable to any compatible agent.

The thesis calls this the outcome economy. A domain expert who packages their judgment into a Skill has produced something that can work for clients without them being present. That is not training. That is manufacturing.

The Challenge​

Create a custom skill for your AI employee that encodes your professional decision-making process. When invoked, this skill should produce output that demonstrates domain-specific reasoning -- the kind of analysis that requires years of experience, not just factual knowledge.

Acceptance Criteria​

  1. A custom SKILL.md exists at .claude/skills/your-skill-name/SKILL.md with valid YAML frontmatter
  2. The skill is invocable via /your-skill-name from WhatsApp
  3. The agent demonstrates domain-specific reasoning when the skill is used (not generic AI output)
  4. At least 5 domain decision rules are encoded in the skill

Here is what a well-crafted domain skill looks like for four professions. Notice how each skill encodes the DECISIONS an expert makes, not just the STEPS they follow.

Accountant: /invoice-reviewer

Reviews incoming invoices against a decision framework: checks for required fields (vendor name, date, line items, tax ID), validates tax calculations against applicable rates, flags amounts that deviate more than 15% from historical averages for that vendor, verifies payment terms match the vendor agreement on file, and rejects any invoice missing a purchase order number for amounts over $500.

Teacher: /lesson-planner

Plans lesson activities using a curriculum alignment framework: maps each activity to a specific learning standard, balances activity types (direct instruction, collaborative, independent practice) across a 45-minute block, checks that prerequisite skills from prior lessons are covered, flags any lesson that introduces more than 3 new concepts for the grade level, and suggests differentiation modifications for advanced and struggling students.

Recruiter: /resume-screener

Screens resumes against job requirements using a structured evaluation: matches required skills against listed experience (not just keyword matching -- looks for demonstrated application), flags employment gaps longer than 6 months for follow-up questions, identifies transferable skills from adjacent industries, scores culture-fit indicators based on the company values document, and generates 3 targeted interview questions based on resume gaps or claims that need verification.

Consultant: /proposal-builder

Structures project proposals using a client-ready framework: breaks the engagement into phased deliverables with dependencies, estimates effort using the firm's standard complexity multipliers, identifies the top 3 project risks with mitigation strategies, includes a "not included" section to prevent scope creep, and flags any deliverable that requires skills outside the team's current capacity.

Hints​

Level 1: Where to Look

Read an existing skill in your NanoClaw installation's .claude/skills/ directory. What YAML fields does the frontmatter contain? How is the instruction section structured? Your skill follows the same format but with your profession's decision logic.

The key fields in the YAML frontmatter are name and description. The description should start with what the skill does and when to use it.

Level 2: What to Ask Your AI

Use this prompt with Claude Code to extract your domain expertise:

I am a [profession] and I need to encode my expertise into an AI skill.

When I [specific task your skill handles], here is my mental process:
- First I check [what you look at first]
- Then I evaluate [what you assess next]
- I flag anything that [your red flag conditions]
- My output includes [what you deliver]

Based on this, what are the 7 most important decision rules I use?
For each rule, express it as: "IF [condition] THEN [action] BECAUSE [reason]"

Focus on the decisions that require EXPERIENCE, not just knowledge.
A new graduate would miss these. An expert would catch them instantly.

Review the output against your actual practice. The AI will generate reasonable rules, but you will notice gaps where your real expertise differs from textbook knowledge. Those gaps are the most valuable rules to add.

PRIMM-AI+ Practice​

This challenge follows the PRIMM-AI+ cycle. Before you build, predict. After you build, investigate.

Predict [AI-FREE]​

Before you start this challenge, write down:

  • What your employee will say when you send a realistic professional task that should trigger this skill, before and after the skill exists.
  • Which 2 of your planned decision rules you expect to produce the most noticeably different output compared to a generic AI response.
  • What the output structure will look like (bullet list, numbered steps, table, paragraph) based on your profession's conventions.
  • Your confidence score from 1 to 5.

Do not ask the agent until those notes are written.

Run​

Use Claude Code ($ claude) to extract your domain decision rules and draft the SKILL.md content. Create the skill at .claude/skills/your-skill-name/SKILL.md, then invoke it from WhatsApp with a realistic professional input.

Complete the challenge as described above: create the skill with valid YAML frontmatter and at least 5 decision rules, invoke it via WhatsApp, and verify domain-specific reasoning appears in the output.

Investigate​

First, write your own explanation of whether the skill output showed genuine professional expertise or generic AI reasoning. Then ask the agent: "Compare my skill's output on this task to what a generic response without the skill would look like. What specific decision rules made the difference?"

Modify​

Change one requirement: replace one decision rule with a stricter version that includes an explicit threshold or condition (for example, change "flag large amounts" to "flag any amount over $500 and state the specific vendor average it deviates from"). Predict how the output will change for a borderline input, update the SKILL.md, re-invoke the skill, and verify the new rule appears in the response.

Make [Mastery Gate]​

Verify against the acceptance criteria above. Passing means: the SKILL.md exists with valid YAML frontmatter at the correct directory path, the skill is invocable via /your-skill-name from WhatsApp, the output demonstrates domain-specific reasoning, and at least 5 decision rules are encoded.

Level 3: Structure Guide

Your skill lives at .claude/skills/your-skill-name/SKILL.md (directory, not flat file). Here is the structure:

YAML Frontmatter (required):

---
name: "your-skill-name"
description: "This skill should be used when [trigger condition]. It [what it does] by [how it works]."
---

Content sections:

  1. Title and purpose -- One sentence: what this skill does and why it matters
  2. When to use -- The trigger conditions (what kind of message or request activates this skill)
  3. Decision rules -- Numbered list of at least 5 rules in the format: check, evaluate, decide, act. These are the core of your skill.
  4. Input format -- What information the skill expects (what the user should provide in their message)
  5. Output format -- What the skill produces (the structure of the response)
  6. Edge cases -- At least 3 situations where the standard rules do not apply and what to do instead

Test the skill by sending a realistic message via WhatsApp that triggers it. Compare the output to what you would produce manually. Where they differ, adjust your decision rules.