Hey Reader,

The email you're reading was drafted by a skill.

I wrote one file, once. It holds my voice rules, my banned words, my structure, the spot where my booking link goes, and two past issues as the standard to match. This week I typed "draft the newsletter on Claude skills" and got a full draft back in my voice. I edited it and hit send.

My LinkedIn posts, my meeting prep, and my weekly briefing run the same way. I run a 12-function marketing operation on Claude for under $300 a month, and skills are the reason it works without me re-explaining my business every morning.

Here's the claim: prompting is labor, skills are infrastructure. A year from now the marketers ahead on AI won't be the best prompters. They'll own the best skill files.

Two things before we start:

1. Save this and block 30 minutes this week to build your first skill.
2. Send it to one marketer who keeps a "best prompts" doc in Notion.

That second one is how this newsletter grows, and I appreciate every forward.

1. What a skill actually is

A skill is a folder Claude reads on its own. Inside is a file called SKILL.md: the instructions, rules, and examples for one recurring job, written once.

At the top of the file sit two lines Claude always sees: a name, and a description of when to use it. Claude scans those descriptions at the start of every conversation. When your request matches one, it opens the full file and follows it. You don't paste anything. The trigger is the work itself.

That's the difference between a skill and the prompt doc you already have. The doc depends on you: dig up the prompt, paste it, then re-explain whatever changed since you saved it. A skill loads itself, every time, with the standard you set on your best day.

One more thing worth having: skills work in the Claude apps, in Claude Code, and on the API. One file follows you across all of them.

2. The rule: third time you paste it, it's a skill

In June I sent you the five-part brief (role, task, material, standard, format). A skill is that brief with the four parts that never change saved permanently. The topic is the only part that still changes, so it's the only part you should still be typing.

The rule I use: the third time I paste the same prompt, it becomes a skill.

What qualifies is any deliverable with a fixed standard that recurs. The weekly email. The LinkedIn post format. The monthly client report. The campaign brief. The QA pass before anything ships.

What doesn't qualify: one-off work and open exploration. A skill is overhead for a job you'll never repeat. Keep prompting those.

3. Write your first one (20 minutes)

Don't write the SKILL.md by hand. Have Claude interview you and write it.

1. Pick one recurring deliverable. The one you re-prompt most.
2. Find two past examples you'd call good. They define your standard better than any adjective.
3. Open Claude and paste this:

Interview me, then write a complete SKILL.md file for a skill that
produces my [deliverable, e.g. weekly client report].
Ask me one question at a time, up to eight questions, covering:
- when exactly the skill should trigger (the phrases I'd type to ask
  for this)
- who the deliverable is for and what they ignore
- my voice: two sentences I'd write, two I'd never write
- the structure the deliverable follows every time
- what a pass looks like, and what I send back for rework
- words and patterns I ban
Then write the file: YAML frontmatter with a name and a trigger
description, followed by instructions specific enough that a new
hire could follow them without asking me anything. I'll paste in
two past examples I rate as good; reference them in the file as
the standard to match.

1. Paste your two examples when it asks.
2. Read the file it writes and cut anything you wouldn't say to a new hire. If a line is generic, it will produce generic work.

The frontmatter description matters more than it looks. It's the only part Claude sees before deciding whether to use the skill, so put your real trigger phrases in it: "Use when I ask for the weekly report, a client update, or paste campaign numbers and ask for a recap."

Screenshot: Claude mid-interview, writing the SKILL.md.

Pro tip: put your banned words in the file. Mine bans "unlock," "elevate," "game-changing," and about thirty others. That one section removes more AI smell from the output than any other instruction I've found.

4. Install it and run it

1. Save the file as SKILL.md inside a folder named after the skill (weekly-report/SKILL.md).
2. In the Claude apps, on a paid plan: Settings, then Capabilities, then Skills. Upload the folder (zip it first if asked).
3. In Claude Code: drop the folder into ~/.claude/skills.
4. Open a new chat and ask for the deliverable the way you normally would. No pasting.

Screenshot: the plain-English request on the left, the on-standard draft on the right.

If it doesn't trigger, the fix is almost always the description. Rewrite it with the exact words you just used to ask.

5. Edit the file, not the chat

This is the habit that separates a skill library from a prompt graveyard.

When output misses, the instinct is to correct it in the conversation: "shorter," "never say that." The output improves, and the correction dies when the chat ends. Next week you make the same correction again.

Move every correction into the SKILL.md instead. Wrong tone? Add the rule to the voice section. Used a word you hate? Add it to the banned list. Each edit takes thirty seconds and stays forever. My newsletter skill has absorbed months of corrections, every one of them made once.

Prompts reset to zero in every new chat. A skill keeps every correction you've ever made. That compounding is the entire argument for building them.

Your first 30 minutes with skills

Minutes 0-5: pick the deliverable and pull two good past examples.
Minutes 5-20: run the interview prompt, paste your examples, trim the file.
Minutes 20-25: upload it (Settings, Capabilities, Skills).
Minutes 25-30: open a fresh chat, ask for the deliverable in your normal words, compare the draft to your past examples.

By minute 30, one piece of your operation no longer needs re-prompting.

Where this falls short

Honesty, as always:

1. Triggering is only as good as the description. Write it vague and the skill never fires, or fires on the wrong task. Use the phrases you'd actually type.
2. A skill scales whatever standard you put in it. Write a mediocre file and you've automated mediocre. Review the first few runs like a manager, then relax.
3. It doesn't know your business changed. New offer or new pricing: edit the file, or it keeps shipping the old ones.
4. One-off work stays faster as a plain brief. Don't skill everything.

One more thing. Everything above is how this operation actually runs, this email included. Anthropic doesn't pay me and I sell no prompt packs. I run marketing on Claude every day and report what holds up.

If you want a second pair of eyes on which of your prompts should become skills, and what goes in the files, grab a free call with me here: https://calendar.app.google/BCm78pVrBnPL3pWf8. It's the working-session version of this guide, on your prompts instead of my examples.

And if this replaces a prompt doc for you, forward it to one marketer who's still pasting.

— Louis

P.S. The single first step: pick the deliverable before you write anything. The best first skill is the prompt you pasted most this month.