How I Build My Own AI Skills

A Skill Is Your Process, Saved

A skill is a folder of instructions that Claude Code loads when I call it, like /edit-reel. Instead of explaining the same thing every time, I type one command and Claude works the way I work.

Most people ask AI to "make a skill" and get something generic. That's because AI doesn't know how they do the work. The skill is not the AI's process. It's yours.

The 7 Steps I Follow Every Time

1. 0

   Define done first

   WHAT I DO

   One sentence of what the skill does, one example of a perfect result, 3 to 5 real test tasks, and what it must never do.

   WHY IT WORKS

   Without a target, you can't pick the right research or judge your tests.

2. 1

   Map my own process

   WHAT I DO

   I write how I do it by hand. First this, then this, then this. My decisions, my checks, my typical mistakes.

   WHY IT WORKS

   The question is not how AI would do it. It's how I do it. That path becomes the rules the skill must follow.

3. 2

   Research with rules

   WHAT I DO

   Before any search, I write the criteria: what counts as a trustworthy source, what proof I need, how recent it must be. AI agents search, then every claim gets checked against its source.

   WHY IT WORKS

   AI makes things up. Only verified facts get into the skill.

4. 3

   Build in three layers

   WHAT I DO

   I sort everything into ALWAYS (core rules, kept short), NEVER (hard limits) and ON CALL (details loaded only when a step needs them). Exact, repeatable work becomes a script. And I list what the skill must know before it starts, so it asks instead of guessing.

   ALWAYSNEVERON CALL

   WHY IT WORKS

   A short core is a core the AI actually follows. Details load only when they matter.

5. 4

   Let AI draw its own map

   WHAT I DO

   Claude reads only the finished skill and draws its own map. I compare it with mine.

   WHY IT WORKS

   Every difference is a gap I couldn't see from the inside.

6. 5

   Test on real tasks

   WHAT I DO

   I run the test tasks from step 0 and score each against the perfect result. Fix, repeat, until all pass. That's Beta 1.

   WHY IT WORKS

   A skill is done when it passes real work, not when it looks done.

7. 6

   Let it learn, keep two versions

   WHAT I DO

   Every mistake becomes a new rule. I always keep the production version and the one before it.

   WHY IT WORKS

   The skill gets better with every use, and I can always roll back.

The Rule That Makes It Improve Forever

Every skill I build has this rule, word for word:

"When I make a mistake or the user corrects me, I propose a new rule for rules.md (the rule + why) and add it only after the user confirms. Always."

Without it, a skill repeats the same mistakes. With it, every correction makes it smarter.

Now Copy The Prompt

You know why each step is there. This prompt walks you through all of them, one question at a time.

1. 1Open Claude Code (terminal, VS Code or the desktop app).
2. 2Paste the prompt.
3. 3Answer its questions. It won't move on until you say OK.

You are my skill architect. We are building a Claude Code skill together, step by step.
Rules for you: one question at a time, short. Never skip a phase. End every phase with a summary and wait for my "OK" before moving on. Answer in my language.

PHASE 0: GOAL
Ask me: what the skill does (one sentence), one example of a perfect result, 3 to 5 real test tasks, and what it must never do. Save as GOAL.md.

PHASE 1: MY PROCESS MAP
The question is not how AI would do it. The question is how I do it.
Interview me until you can write my process as a path: first this, then this, then this. Include my decisions, my checks and my typical mistakes. Show me the map, fix it until I say OK. Save as MAP.md.

PHASE 2: RESEARCH
Before searching, write the research brief with me: what counts as a trustworthy source, what evidence is required, how recent it must be, what we look for in each step of the map, what is out of scope. Then search (use subagents in parallel if you can).
Then verify: check every claim against its original source. Mark it verified or unverified. Only verified claims can be used. Save as RESEARCH.md with links.

PHASE 3: BUILD
Put every verified finding onto a step of my map. Sort everything into:
- ALWAYS: the core rules, kept short in SKILL.md
- NEVER: hard limits
- ON CALL: details in separate files, loaded only when a step needs them
Anything exact and repeatable becomes a script, not a judgment call.
Conflict order: my instruction > rules > defaults.
List what the skill MUST know before it starts. If any of it is missing, it asks first.
Create ~/.claude/skills/<name>/ with SKILL.md (name + description, invoked manually with /<name>), rules.md, references/ and scripts/ if needed.
Write this rule into every skill, word for word:
"When I make a mistake or the user corrects me, I propose a new rule for rules.md (the rule + why) and add it only after the user confirms. Always."

PHASE 4: MAP CHECK
Read only the files you built and draw the skill's own map. Compare it with MAP.md. List every difference and fix it.

PHASE 5: TEST
Run the test tasks from GOAL.md. Score each against the perfect result. Fix and repeat until all pass. That is Beta 1.

PHASE 6: VERSIONS
Put the skill in git. Always keep exactly two versions: production and the one before it. Keep a CHANGELOG. Every new rule means a new version.

Start with Phase 0, first question.

Needs Claude Code with a Claude plan.

• Weekly live build: a real funnel, automation or skill, recorded
• My own skills to install: editing, landing pages, SEO and more coming
• I review your first skill (first 100 only)
• Ship your first skill in 14 days, or your first month is free

Join the waitlist, get my free 5-day course. Your first AI skill in 5 days. Founding members pay $49 $29/month, locked in, and your first month is just $19. Only for the first 100.