Create a working skill in ~10 minutes
Let’s stop talking about skills and build one. Right now. Ten minutes.
Choose how you want to build:
Copy this prompt into your AI tool. It will ask if you already have a task in mind, or help you find one. Then it walks you through building a complete SKILL.md file.
I'm taking a course on Agentic skills — reusable instructions that teach AI how to do a specific task my way, using the agentskills.io standard. I need to build my first skill. Start by asking me: "Do you already have a task in mind, or would you like help choosing one?" If I have a task, ask me to paste or describe it so you understand what I'm working with. Then move on to the interview. If I don't have a task yet, help me find the right one by asking: 1. What tasks do I currently use AI for on a regular basis? 2. Which of those frustrates me most or takes the most time when the output isn't right? 3. For that task, do I end up re-explaining the same rules every time? Help me narrow it to something specific if it's too broad (like "help me with emails" → "write a client follow-up email in my tone"). Once we have the task, interview me one question at a time: 1. When should this skill activate? What phrases would I say to trigger it? 2. What are the steps to do this task, in order? 3. What are the rules? (always do X, never do Y) 4. What does great output look like? Can I paste an example? After the interview, generate a complete SKILL.md file using this format: --- name: [kebab-case-name] description: "[What it does]. Use when [trigger phrases]. Do NOT use for [what this isn't]." --- # [Skill Name] ## Steps [numbered steps] ## Rules [bullet list] ## Example [concrete example of good output]
Save the SKILL.md your AI generated, then install it:
What AI tool do you use?
AI tools are getting better at this. Claude’s skill-creator now runs a structured interview to scaffold your SKILL.md (what should it do, when should it trigger, expected output format, should we set up tests). OpenAI’s Codex CLI has a similar workflow. The manual approach taught in this lesson is still the right starting point, because understanding the structure matters before you let tools automate it.
Takeaway: You now have a working skill installed in your AI tool. It’s probably not perfect — that’s what the next lesson is for.
Progressive disclosure — skills load in 3 levels:
Name + description (~100 tokens) — loaded at startup for all skills. Like a job posting on a bulletin board. This is how the AI decides whether to activate your skill.
The full SKILL.md body (< 5,000 tokens recommended) — loaded when activated. Like handing someone the training manual for a specific task.
References, assets, scripts (variable) — loaded only when needed. No context cost until accessed. Covered in Lesson 4.
Why this matters:
AI has a limited “attention window” — everything in the conversation takes up space. Pasting a 2,000-word prompt into every chat burns through that window fast. Skills load only the relevant instructions, saving 50%+ of token usage.
Full SKILL.md anatomy:
--- lines) — name (kebab-case), description (when to activate)Naming rules:
weekly-status-report)SKILL.md (case-sensitive)Structuring tips:
Open standard: Skills follow the SKILL.md format from agentskills.io, supported by 35+ tools. They’re portable — write once, use everywhere.
We work alongside your team to build AI-native workflows — from one-week sprints to full engineering acceleration. No handoffs, no slide decks.
Talk to us