If you explain the same thing to Claude twice, that's a skill waiting to be saved.
A skill is just a folder with instructions. That's it. Not a plugin, not an app, not something you need to code. A folder with a markdown file that tells Claude how to do something you do all the time.
You're going to identify one workflow you repeat constantly, package it as a skill Claude can run on command, and set up a keyboard shortcut that saves you 30 seconds every time you use it. Multiply that by 10 uses a day.
Think about the last week. What did you explain to Claude more than once? What do you do 5+ times a week that follows the same pattern?
Examples for creative directors:
Pick one. The simplest one. That's your first skill.
mkdir -p ~/strategy/skills/[your-skill-name]
A skill has one required file: SKILL.md. That's the instruction set. Claude reads it when you invoke the skill.
Tell Claude what you want:
I have a workflow I repeat all the time: [describe it].
Create a skill for it at skills/[name]/SKILL.md.
The SKILL.md should include:
- When to use this skill
- Step-by-step instructions for Claude to follow
- What the output should look like
- Any preferences about tone, format, or structure
Use my words, not generic language.
Claude writes the instruction file. Review it. Edit it. It should sound like you giving directions to a sharp assistant.
Open Claude in your strategy folder and reference the skill:
Use the skill at skills/[name]/SKILL.md to [do the thing]
Watch Claude follow your instructions. If the output is 70% right, tell it what to change. Refine the SKILL.md until it nails your workflow.
Here's a real example. This skill turns rough meeting notes into structured action items, sorted by person, with deadlines pulled from context:
# Meeting Notes Processor
## When to use
After any meeting where I have rough notes or a transcript.
## Instructions
1. Read the notes or transcript
2. Extract every action item mentioned
3. Group by person responsible
4. Add deadlines if mentioned, flag "no deadline" if not
5. Format as a checklist I can paste into Slack
## Output format
- **[Person Name]**
- [ ] Action item (by [date] or "no deadline")
## Preferences
- Be direct. No filler.
- If something is ambiguous, flag it, don't guess.
- Keep it short. I'll add context if needed.
You already have Shift+Cmd+Q from Lesson 01. But there's more you can do.
The idea: any action you take 10+ times a day should be one keystroke. Even saving 3 seconds per use adds up. At 15 uses a day, that's 45 seconds. Over a week, 5 minutes. Over a year, 4+ hours. From one shortcut.
To add more Hammerspoon shortcuts, open your config:
-- Example: Shift+Cmd+N to open a new capture in your brain inbox
hs.hotkey.bind({"cmd", "shift"}, "N", function()
hs.applescript([[
tell application "Terminal"
activate
do script "cd ~/strategy && claude 'Create a new file in brain/00-inbox/ for a quick capture. Ask me what I want to capture.'"
end tell
]])
end)
Skills are the application layer for your thinking. Your CLAUDE.md teaches Claude who you are. Skills teach Claude how you work. Together, they mean Claude gets better at your specific workflows every time you use it.
Most skills automate a task. The interesting ones change Claude's behavior permanently. Here are six that form a system. Together they prevent mistakes, correct the ones that happen, capture everything, and let you step away.
| Phase | Skill | What it does |
|---|---|---|
| Prevent | /preflight |
Auto-selects the right workflow before execution starts. Prevents the wrong work from beginning. |
| Prevent | /calibrate |
Asks clarifying questions until 95% confident. Kills rework at the source. |
| Correct | /principal |
Diagnoses the principle underneath a mistake, designs a structured encounter, saves it. Corrections become permanent learning. |
| Correct | /teach |
Converts "don't do that" into Socratic questions. Transfers judgment through questions, not orders. |
| Capture | /learn |
End of session, Claude reviews the conversation and updates its own CLAUDE.md. The knowledge grows automatically. |
| Trust | /walk-away |
Generates an autonomous operating brief. You leave, Claude keeps shipping. Trust as a product. |
The loop: /preflight and /calibrate prevent mistakes before they start. /principal and /teach correct the ones that get through. /learn captures everything. And once enough lessons compound, /walk-away becomes possible. You stop supervising and start directing.
When Claude makes a mistake, instead of just correcting it, this skill diagnoses the principle underneath the surface rule, designs a structured encounter with that principle, and saves it to a knowledge system. Four steps, 82 lines:
# Step 1: Diagnose the principle (not just the surface rule)
SURFACE RULE "Don't refactor adjacent code when fixing a bug"
PRINCIPLE Scope Discipline: the requester defines the boundary.
Proximity is not permission.
# Step 2: Design the encounter (concrete, not abstract)
ENCOUNTER: scope-discipline
VIOLATION Refactored working drag-and-drop while fixing a bug
PRINCIPLE The requester defines the boundary
COUNTER-CASE Should have: read the bug, fixed only that code path,
tested drag-and-drop still works, committed a 4-line diff
RULE If the user didn't ask you to touch it, don't touch it
# Step 3: Persist (so it compounds across sessions)
# Step 4: "Back to your bug."
What makes this different from a regular skill:
The first time we tested /principal, we ran it on its own first draft. The skill about principles-over-rules had shipped with a rule list. It caught its own contradiction. That's when you know a skill is working.
You don't need to build all of these today. But know that skills can go beyond "do this task for me" into a system where Claude prevents mistakes, learns from the ones that get through, and eventually earns enough trust that you can walk away. That's where the real leverage lives.
Build one more skill this week. Pick the workflow you repeat second-most. After a week with two skills running, you'll start seeing the third, fourth, fifth ones everywhere. That's the pattern recognition kicking in.