1 / 67
Code for Creatives

Week 2.

Your First Skill.

April 20, 2026
Notes ▲

you are not behind.

last week you wrote a CLAUDE.md. that's real. celebrate it.

Notes ▲
Open warm. Whatever they did last week is enough to start week 2 with. Permission slip.

today!

Notes ▲

quick check-in.

show of hands (or drop in chat):

Notes ▲
Don't skip. Calibrates the room and reminds people they did something real.

context is for this week.

we left week 1 with a promise. here it is.

Notes ▲

three quick tips first.

things you'll actually use this week.

Notes ▲
tip 01

drag images into the terminal.

Claude can see them. screenshots, photos, mockups.

try: paste a screenshot, say "what do you see? what would make this better?"

Notes ▲
tip 02

clipboard is a tool.

paste to my clipboard

don't want weirdly formatted text? have claude paste to clipboard

Notes ▲
tip 03

know your limits.

type /status in claude code. see your session usage.

you have more runway than you think. but don't let a runaway agent eat your week.

Notes ▲
Mention the Jeffrey story — accidentally ran a skill-creator eval loop, burned his usage. /status tells you where you stand.

what is a skill?

Notes ▲

it's just a file.

a markdown file. english. telling Claude how to do a specific job.

> ls ~/.claude/skills/
# chief-of-staff/
# c4c/
# meal-planner/
> cat ~/.claude/skills/chief-of-staff/SKILL.md
# "you are alex's chief of staff..."
Notes ▲
No code. No API. Text files in a folder. Read the SKILL.md line out loud. Feel how mundane that is. That's the magic.

CLAUDE.md vs. skills.

same language. different scope.

Notes ▲

how a skill fires.

two ways in:

Notes ▲
The description field is a TRIGGER, not a summary. Write it as "when should I fire?" — "fire when alex mentions writing" — not "what do I do." Claude picks skills by reading descriptions. This is the single highest-leverage thing they can learn today.

let me show you one.

/chief-of-staff
Notes ▲

this is the real one.

i use it every morning. let's take it apart.

Notes ▲
Deep Dive

The Chief of Staff.

How I built a thinking partner

Code for Creatives / Session 2
Notes ▲
This is the real skill I use every day. Not a demo, not a toy. Let's take it apart and see what makes it work.
One command

/chief-of-staff

Claude transforms from a blank assistant into a thinking partner that knows my whole world.

Notes ▲
Type this live. Show them what happens. The skill file loads, context floods in, and Claude becomes a different thing.
Under the hood

Four things happen.

every session, every time, automatically

Notes ▲
Reveal one at a time. Each step is written in English in the skill file. None of this is code.
Step 1: Load context

It reads everything it needs.

every file here is just text. no database. no API. just markdown.

Notes ▲
Point at each one. The magic is in the reading, not the format. You tell it which files to read in the skill file. In English.
Step 3: Building the brief

How it assembles.

Claude reads all those inputs, then cross-references them:

then it writes four sections. same format every time.

Notes ▲
Reveal one at a time. Each bullet is a cross-reference Claude does. The skill file just says "read these files, pull the calendar, build a brief." Claude figures out the connections.
This morning's actual output

The real brief. Today.

TODAY'S PRIORITIES
1. Content HQ demo (OAT / Britt) - happening TODAY
"Don't walk in without the data prepped."
2. Upload Matt Abrahams Ep7 - DUE TODAY
Episode drops TOMORROW. Upload window closes today.
3. Cohort 2 Week 2 reminder email - DUE TODAY
"Write it now. It's a 10-minute task."
ONE THING WORTH FLAGGING
The Fran story is a proof-of-concept asset OAT doesn't fully see yet.
WHAT CLAUDE CAN DO RIGHT NOW
- Draft the Week 2 reminder email
- Draft Fran mini-interview questions
- Build Charlotte expansion content brief
Notes ▲
This is the actual output from this morning. Real tasks, real deadlines, real pushback. Notice the tone. That's the values and personality showing up.
From the same brief

Meeting takeaways.

4 meetings. Auto-transcribed. Summarized. Action items pulled. No notes app needed.

OAT Call - Fran / Charlotte Expansion
- Fran's story is the story.
- Action (Alex): Set up mini interview with Fran.
OAT All-Hands - March Strategy
- Britt's own voice wins on LinkedIn.
- Cold outreach is breaking.
Channel Strategy Sub-meeting
- No decision made. YouTube + events are agreed path.
Transcript 4 (Thinking Stories)
- Stray recording. Lauren's course. Not yours. No action.
Notes ▲
Point out the last one. Claude figured out that one of the 4 transcripts wasn't even my meeting. It dismissed it. That's context at work.
The secret ingredient

Values, not vibes.

Every decision gets filtered through four anchors:

Notes ▲
This is the part people don't expect. The skill pushes back. "Is this aligned with your truth?" That's the difference between a to-do list and a thinking partner.
It asks me things like

The questions.

"Is this something you want, need, or think you should do?"
"What's the real deadline, and what happens if it moves?"
"Who is waiting on you and for what?"
"What would happen if you just... didn't do this?"
Notes ▲
These are all written in the skill file. In English. "When I'm spiraling, ask me these questions." You're programming Claude's behavior with sentences.
The memory loop

How it remembers.

Every session writes back to the same files it reads from:

it reads the files. it updates the files. next session, it reads them again.

Notes ▲
This is the key insight. Claude doesn't have built-in memory between sessions. But text files persist. Read at start, write at end. That's memory. It's just files.
The vibe

It has a personality.

That's literally in the skill file. English instructions for how to sound.

# Tone
Blend of Seth Godin and Robin Sloan.
Thoughtful, curious, creative, grounded,
slightly quirky, unfailingly honest,
always returning to the essential thing.
Notes ▲
You can tell Claude what KIND of thinking partner you want. Not just what it does, but how it feels. Two writers I admire, mashed together. That's a skill instruction, not code.
Timeline

How it grew.

Notes ▲
It didn't start as this complex system. It started with maybe 20 lines. Every addition came from "I wish it also did X." That's what they're doing this week with their first skill.

Not a chatbot.

Infrastructure.

Notes ▲
Let this sit. The shift from "I ask Claude things" to "Claude is running my operations in the background" is the whole arc of this course.
Your turn

Start with one question.

Your version just needs to answer one question you ask yourself every day:

Notes ▲
Bring it back to them. This isn't about my system. It's about what question they want answered every morning. That's their v1.
The pattern

Every skill grows the same way.

repeat forever

Notes ▲
This is the whole course in one loop. Write, use, fix, grow, repeat. The Chief of Staff is what happens when you do this for a few months.
Label

You can invoke a skill

Or claude can invoke it for you. 

Notes ▲

how does claude know to run a skill?

the description is the trigger.

the description is not a summary. it's a sentence claude will use to decide.

Notes ▲
THIS is the whole game. Write the description like a trigger — not like a summary. 'Helps with writing' is worthless. 'Fire when alex mentions drafting, revising, or opens a blog file' is a switch claude can flip.

skills vs. agents.

the terminology is confusing. there are really just two things.

Notes ▲

skills are rooms you've decorated.

when you type /chief-of-staff, you walk into a room.

Notes ▲
The spatial metaphor is the point. You're not using a tool. You're entering a space.

agents are runners you dispatch.

you send them out. they work. they come back. you stay in your room.

Notes ▲
You don't type these. Claude sees your ask, decides 'this needs code-builder' or 'this needs Explore', spawns one with instructions, reads the result, continues. You stay in your conversation the whole time.

the difference.

one phrase each:

Notes ▲

rooms i've built.

each is a context. a vibe. a set of files to read.

Notes ▲
Each is a room I walk into. Same terminal, different world. The skill file decides what Claude reads, what tone it uses, what it focuses on.

agents claude dispatches for me.

i never type these. claude decides to spawn them:

more invocations than every skill combined. most of my real compute goes through agents.

Notes ▲
Drive home: you don't need to know code-builder exists. Claude reaches for it when the task calls for it. The agent's description does the picking — same pattern as skills. Skills you type (sometimes). Agents you never type.

one more, in progress this week.

britt started a youtube channel. 10 videos coming. ~20 deliverables per video. i am not doing that by hand 10 times.

Notes ▲
Transition from cos_growth: 'so that's /chief-of-staff — a few years deep. let me show you one that's two weeks old and still being built.' This lands the point: skills are LIVING, not finished.

what /oat-youtube actually does.

one command. 20+ deliverables. three stages.

stage 1 — understand
01ingest raw video
02transcribe (via /transcribe)
03analyze → section map
stage 2 — generate
04render overlay clips
05render opening animations
06placement guide (timecodes)
07thumbnail variants
stage 3 — package
08titles + descriptions
09package to clipboard
10update pipeline tracker

each step is a paragraph in SKILL.md. the skill IS the flow.

Notes ▲
Walk the stages from left to right. The point: the skill isn't code. It's a numbered list of "do this, then this, then this" — plus the constants that are easy to forget (frame rate, color tokens, pixel format). A subagent can follow it. A new teammate could follow it. Future-me can follow it. The skill = the pipeline.

the signal: i'm going to do this again.

that's when you write the skill. not before.

Notes ▲
THE lesson. Do the work once by hand. Then the moment you realize you're about to do it a second time — THAT's when the skill gets born. Not when the idea sounds fun. V1 of the skill is informed by having done the thing.

every failure becomes a rule.

the skill gets sharper every week. that's the whole job.

Notes ▲
All three are real bugs from this past week on britt's episode 2. Each one became a line in the skill. The point: don't fix the problem in your head — fix it in the file. Next time is free.

ROI shows up on video two.

video one is how you learn what the skill even needs.

Notes ▲
Don't try to break even on your first use. You won't. Skills pay off the second, third, tenth time. The homework is built around this — start using it, let the friction teach you what to change.

a skill that's four lines.

/grill-me. i use it 16 times a quarter. here's the entire file:

---
name: grill-me
description: Interview the user relentlessly about a plan or design until reaching shared understanding. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
---
Interview me relentlessly about every aspect of this plan until we reach a shared understanding.
Walk down each branch of the design tree, resolving dependencies one-by-one.
Ask the questions one at a time.
If a question can be answered by exploring the codebase, explore the codebase instead.

that's the whole skill. no scripts. no tools. just instructions. 94% of the time claude picks it on its own.

Notes ▲
This is the opposite of /chief-of-staff. COS is a room I decorated for years. /grill-me is 4 instructions and a tight description. Both are real. Both get used. The SIZE of the skill has nothing to do with its usefulness. Some of my best ones are 4 lines. Don't wait until you have a masterpiece.

the skills i actually use.

last 90 days, top 10 by count:

472 fires across 87 different skills in 90 days.

Notes ▲
Real numbers from my logs. Most are 3-10 line markdown files. The most-used ones are always the ones I wrote for a specific repeated job — not the clever ones I thought would be useful.

zero-to-one happens fast.

you can go from barely using skills to skills being half your workflow in 6-8 weeks.

Notes ▲
Push this one. The jump from 4 to 204 in one month happened because I stopped treating skills as nice-to-haves and started treating them as infrastructure. They're at month 0 right now. What they build in the next 6 weeks compounds for years.

and you already started. (week 1 builds)

from last week's show-and-tell + the wiki:

every one is a first build. zero are polished. all of them work.

Notes ▲
Name these four by name. Look at chat when you say each name. The point: none of these are week-3 masterpieces. They're week-1 fireballs. If anyone feels behind, you're not — most of these people also felt behind, and then they built.

you don't have to build them all.

skills.sh is a package manager. search, install, use.

Notes ▲
People think they have to write every skill from scratch. They don't. There's a whole ecosystem of shareable skills — and a CLI to find + install them.

npx skills find [query]

search interactively or by keyword:

skill discovery IS a skill. fire /find-skills and describe what you wish you had.

Notes ▲
Demo if time: just say 'find me a skill for writing tests' and watch claude fire /find-skills. The answer is usually 'there's already one at github.com/foo/bar, want me to install it?' — and then you can.

but the best skills are the ones you make yourself.

only you know exactly what you keep doing. only you know your voice.

Notes ▲
Counterpoint to skills.sh. The ecosystem is great for niche tooling (design systems, test frameworks, language parsers). But the skills you fire 50x a month are always the ones YOU wrote for YOUR specific work. No catalog will ever have /warm-up-for-alex or /vibe-coach-for-you. Write it yourself — in english — and it becomes yours.

a morning brief that reads your actual inbox.

the skill file points Claude at the Gmail MCP. Claude pulls unread, triages, drafts replies.

# chief of staff / inbox scan
# newer_than:1d is:unread
> needs your brain (3): britt re: scope doc, connor podcast, CMI invoice
> i handled (11): newsletters, confirmations, receipts
> nothing on fire. you're clear until 2pm.
Notes ▲

watch me build one with you. right now.

60 seconds. no magic. just a file, a name, a description.

Notes ▲
Plan: build /warm-up live. Open terminal. mkdir ~/.claude/skills/warm-up. touch SKILL.md. Write 5 lines (name, description, body). Save. Type /warm-up. It fires. Total: ~60 seconds. If it fails, that's the lesson — check the description. This shows skill creation is NOT a coding exercise.

BUILD TIME.

25 minutes. build your first skill.

screens off. hands on. we'll regroup for show + tell.

Notes ▲

your annoying thing is your candidate.

last week's homework asked you to find one. that's your first skill.

Notes ▲

skeleton of a skill file.

starter template. copy, paste, change the names:

---
name: morning-writer
description: fire this in the morning to help alex start his writing session. asks one question and gets him moving.
---
# you are alex's morning writing coach.
## first, check the date and time.
## then, read ~/writing/current-draft.md
## then, ask one question: "what's the next sentence?"
## do not offer 3 options. do not frame it. just one question.
Notes ▲
The description matters MORE than the body. Write the description as a trigger sentence. If it's YAML, keep it on ONE LINE. Multi-line descriptions silently disappear — a real gotcha people are hitting in the last month.

stuck? here are 10 to steal.

narrow. obvious. useful tomorrow.

pick one. write it in english. save it. use it.

Notes ▲
Read 3-4 of these out loud. Let them echo. The students who can't think of one are hearing 'wait, i want that.' That's the candidate. Don't pitch all 10 — pick the ones that land for the room energy.

show + tell.


Notes ▲

broken is fine.

tried is the whole point.

Notes ▲
homework
before next week.
  • 01 use your skill every day. notice what's wrong. fix it.
  • 02 teach Claude to improve your skill: "here's what it gets wrong, rewrite it."
  • 03 drop a screenshot in #show-and-tell. the ugly version is the point.
  • 04 (optional) plug in Gmail or Calendar MCP. make your skill touch real stuff.
full doc lives in slack with copy-paste prompts for each.
Notes ▲

want to go further?

optional advanced track:

Notes ▲

you're building your memory.

every skill you write is a piece of how you work, captured.

Notes ▲

quick taste. what CLAUDE.md can become.

we're not touching this today. but here's what it grows into.

Notes ▲
Speed through this block. The point is aspiration, not detail. You want them walking away thinking 'wait, I can do THAT?' — not taking notes.

twelve things living in my setup right now.

CLAUDE.md isn't configuration. it's a living relationship. every friction becomes a rule.

Notes ▲
Read the subExtra at the end. THAT is the punchline. If they ask 'how do I set up X?' — say 'we'll get to a lot of these in the coming weeks.' Don't rabbit-hole.

What if Claude already knew your priorities?

Notes ▲
Pause. Let them sit with it. Most people start fresh every conversation. What changes when Claude has memory?

zoom out.

/chief-of-staff is one. let me show you the rest.

runners i dispatch.

these go do work and report back:

the crown wears a cron.

/chief-of-staff fires 54 times. i type it maybe twice. the other 52 are my mac mini running it while i sleep.

Notes ▲
The distinction: some skills you fire, some skills fire at you. COS is the second kind — morning brief hits my phone by 6am because a cron job fires the skill. My real hand-invoked king is /here-now at 46. Teaching point: some of your best skills will be ones you never type. They type themselves.

a room is a commitment. a runner is a verb.

i built rooms first. 90 days later, 7 of my top 10 are runners. build more verbs.

Notes ▲
Rooms are full worlds — heavy to build, only useful in one context. Runners are single-purpose tools that compose. /humanizer lands in every pipeline. /here-now publishes any page. Build narrow runners and they'll find each other in sessions.

integrations = more files it can read.

MCPs are small servers that give Claude new tools. install once, reference in your skill.

Notes ▲
This is the "going further" lane. Don't sell it as the baseline. Most students will get value from a pure text-file skill first. But for the ones who are ready: this is what unlocks the real chief-of-staff feeling.

see you wednesday.

same time. bring what you broke.

you can plug it into real tools.

gmail. calendar. slack. notion. anything with an API.