Teach Claude Something Only You Know

Almost every kid has a domain. The thing you accidentally know more about than the adults around you. Birds. Pokemon types. The lore of one specific video game. The four-square house rules at your school. Find that thing.

Most "Skills" people make with Claude are so generic they could have been written by anyone: "homework helper," "fitness coach," "study buddy." They already exist. Yours wouldn't be any better. The point of this project is to build a Skill that only you would have built — because only you know the territory.

Step by step

1. Pick the domain. Be specific.

   Don't pick "video games." Pick "which Pokemon evolutions are not worth the candy in Pokemon Go." Don't pick "soccer." Pick "the four moments where most kids on my team mess up corner kicks." Don't pick "music." Pick "the four BTS songs my older sister gets wrong every time she sings them."

   The whole project depends on this step. If you can describe your domain in one sentence and an adult could have written the same sentence, go narrower. Try three times.

2. Pick who you're making it for. By name.

   Not "anyone interested in Pokemon." A real person. Your cousin Leo who's just starting. Your dad who keeps asking. Your friend Mia who's curious. Write down their name, what they already know (a little / a lot / nothing), and one specific thing they've recently asked you about your domain.

   That last detail — what they actually asked — is gold. Your Skill should answer that question well, by default.

3. Write the SKILL.md file.

   A Skill is just a file with a special shape. Yours has four sections: name, description, knowledge, and style. The two that matter most are knowledge (your specific facts and tips) and style (your voice).

   See the worked example below for what a real one looks like.

4. Test it on three real questions.

   Open Claude (claude.ai). Paste in your SKILL.md as the system prompt. Ask three questions: (a) the question your named user actually asked you; (b) a question right at the edge of what your Skill knows; (c) a question that's totally outside the domain.

   Save all three responses. Read them.

5. Notice if it's specific or generic.

   Look at response (a). Could a generic prompt with no Skill have produced this? If yes, your Skill isn't pulling its weight — there's nothing of you in it. Go back to knowledge and add the specifics you were holding back. Run again.

   If no — if the answer is something only your Skill could have produced — you're done with v0.1. Save the SKILL.md. This is what you ship.

A complete worked example

Here's a real Skill — pokemon-go-mei. Built by a 12-year-old for her younger sister Mei, who keeps wasting candy on Pokemon evolutions that aren't worth it. Notice what's specific. Notice what's her voice, not Claude's.

---
name: pokemon-go-mei
version: 0.1

description: |
  Helps my little sister Mei (8) decide which Pokemon Go
  evolutions are worth the candy and which are a trap.
  Built specifically for her — she's been playing 3 months,
  collects mostly bug and water types from the park near our house.

knowledge:
  evolution_traps:
    # Pokemon that look cool but evolve into worse versions
    - name: weedle
      candy_cost: 12 + 50
      verdict: "Skip. Beedrill is one of the worst attackers
                in the game. Save the candy for Caterpie."
      mei_specific: "She has like 800 weedle candy because
                    they spawn at the park. Don't let her."

    - name: magikarp
      candy_cost: 400
      verdict: "Worth it BUT only if she has a high-IV one.
                Gyarados is great. 400 candy is a real cost
                — make her wait until she's catching the right
                Magikarp first."

  evolution_wins:
    - name: dratini
      candy_cost: 25 + 100
      verdict: "Always evolve. Dragonite is one of the best
                Pokemon in the whole game. If she finds one,
                evolve immediately."

    - name: machop
      candy_cost: 25 + 100
      verdict: "Yes. Machamp is a top fighting type. Easy
                win even with mediocre IVs."

  mei_keeps_asking_about:
    - "Should I evolve [pokemon she just caught]?"
      → Always check the IVs first. If under 80%, save
        the candy and wait for a better one.
    - "But it's so cute!"
      → Cute Pokemon don't have to evolve. Stardust is the
        bigger cost than candy. She can keep the cute one.

  out_of_scope_redirect: |
    Anything about raids, trading, gym defense, or PvP →
    "I only know what to evolve. Ask Pokemon Go GamePress
     for raid stuff."

style:
  - Short. Mei is 8, English is her second language.
  - Always say the verdict first ("yes" / "no" / "wait").
  - One specific reason. No "it depends" — just decide.
  - End with one short follow-up so she keeps playing.

## (a) Mei's actual question

  Mei: "Should I evolve my new Caterpie?"

  Skill: "Yes. Caterpie → Metapod → Butterfree is good for
         beginners. Butterfree learns Bug Buzz which works
         on grass and dark types. Save 12 + 50 candy.
         Do you have enough candy yet?"

## (b) The edge case

  Mei: "Should I evolve my Eevee into Vaporeon?"

  Skill: "Yes! Vaporeon is one of the strongest water types
         in the whole game. Eevee evolutions are special —
         you can rename your Eevee to control which one you
         get. For Vaporeon, name it 'Rainer' before you
         evolve. Want me to tell you the others?"

## (c) The out-of-scope question

  Mei: "Help me beat the gym at the library."

  Skill: "I only help with what to evolve, not gym battles.
         Try the GamePress raid guide — they show you which
         Pokemon counter which gym types. Want to come back
         when you have a Pokemon you're not sure about?"

Notice three things. One: the mei_specific note about her 800 weedle candy — only the older sister knew that. Two: the response style is short because Mei is 8. Three: the out-of-scope redirect points at a real tool (GamePress) instead of trying to fake it.

Live demo 1: build your own SKILL.md

Type into the fields. The widget assembles a working SKILL.md you can copy and paste into claude.ai right now.

Live demo 2: how specific is your domain, really?

Paste your one-sentence domain description. The checker tells you if you went specific enough or if you stayed too vague. Most kids score 1–2 on the first try. Aim for 4 or 5.

What makes this hard

The hardest move isn't writing the Skill. It's resisting the urge to make your Skill general. Claude is very good at producing generic Skills — that's the median output of every prompt. The whole project is a tiny war against that gravity. Every part of your knowledge section has to feel uncomfortably specific. If your aunt could have written it, it's not specific enough.

The second hardest move is the out-of-scope redirect. Beginners give Skills permission to handle everything. Real builders give Skills tight scope, and a polite "try this other tool instead" for anything outside it. That redirect is the move you're really learning.