Claude Skills: A simple way to stop repeating yourself
I was over coffee this week with a guy who runs a small business here in town — not in publishing, different industry entirely — and he was telling me how much time he spends every morning writing customer emails.
"I open ChatGPT, I paste in the same paragraph about how my business sounds, I describe the situation, and then I can finally get help drafting the reply. By the time I'm done explaining, I could've just written the email myself."
I told him there's a fix for this in Claude, and it takes about 30 minutes to set up. It's called a Skill — a reusable set of instructions Claude loads automatically whenever it's relevant. Build it once, never paste those instructions again.
This applies to anyone who uses an LLM for the same kind of task more than twice a week. Here's how I walked him through it.
What you'll need
A Claude account (the free tier works for this)
A short list of "how I do this" preferences — even just bullets in a notebook
One real example of the kind of task you want help with, for testing
No coding, no plugins, no setup beyond clicking around in Claude's settings.
Step 1 — Dump your preferences (10 minutes)
Open a blank doc and write down what you'd normally have to tell the AI before it could help you. Don't try to be exhaustive. Just dump what you can think of in 10 minutes.
For my friend's customer-email Skill, his list looked something like:
Tone is friendly but professional — no "Hi there!" or exclamation marks
Always sign off with "Talk soon, [name]"
If a customer is upset, acknowledge the frustration before offering a solution — never lead with the fix
Keep responses to 3 short paragraphs max
Don't make promises about delivery times
If the question is technical, suggest a phone call instead of trying to solve it over email
About 12 bullets. If you're stuck, look at the last few responses you wrote yourself and notice what makes them sound like you and not a chatbot.
Step 2 — Ask Claude to build the Skill (5 minutes)
Open a new Claude chat and paste this prompt, with your bullets at the bottom:
I want to create a Skill that captures how I write customer emails so I don't have to paste these instructions into every chat.
My preferences:
[paste your bullets here]
Please draft a Skill I can save. Keep it under 50 lines. The goal is for Claude to apply these rules automatically when I ask for help replying to a customer.
Claude hands you back a tidy Skill, usually 30 to 50 lines, organized into sections. Read it once. If something's off, tell Claude what to change.
Step 3 — Save it (2 minutes)
In Claude, go to Settings → Capabilities → Skills, click "New Skill," paste in what Claude generated, and give it a name. Save it.
Step 4 — Test it on real work (10 minutes)
Most people skip this step. It's the most important one.
Open a fresh chat. Don't mention the Skill — Claude should know to use it on its own. Paste in a real customer email you got recently and ask Claude to draft a reply.
If the Skill is working, the output should already sound like you. If it doesn't — if you find yourself making the same correction twice — that correction belongs back in the Skill. Edit and re-save.
What worked
The first reply my friend ran through it came back about 90% of the way there. He changed a couple of words. Compared to his usual "delete the whole thing and start over" pass, that's real time saved.
The biggest win was tone consistency. He said his replies had been drifting more formal over the years and his Skill pulled them back to the warmer register he'd built his business on.
What didn't
The Skill was too eager to apologize. A customer complained about something that wasn't actually his fault, and the draft opened with "I'm so sorry for the trouble" — which would've conceded a point he didn't want to concede. He had to add a rule: "Acknowledge the frustration without taking blame for things outside our control."
Lesson: any rule that runs contrary to what an AI would naturally do — don't apologize unprompted, don't soften a hard message, don't smooth out a deliberate edge — has to be stated explicitly. Claude defaults to friendly.
What I'd do differently
Start smaller. His first version had 12 rules. Half of them never came up in practice. If he were starting over, I told him, build a Skill with just 4 rules — the ones you correct most often — and let it grow as you use it.
Your turn
This works for any task you do repeatedly with the same set of preferences. Customer emails, proposal templates, internal team updates, social media posts, vendor follow-ups — anywhere you find yourself pasting the same instructions twice, you have a Skill candidate.
If you build one this weekend, hit reply and tell me what you put in it. I'll share the best ones (with permission) in a future Friday.
