Last week I turned three of my friends into domestic invalids.
Showed them Claude Code. An agent that lives in your terminal. Sees your files, reads them, writes new ones, runs commands. You give it a task — it breaks it into pieces, spawns a sub-agent for each piece, checks the result.
They haven’t responded to messages in a week. When they do surface, they sit there like children. Staring at the terminal. Giggling. Ignoring their employees.
All three are serious people. Entrepreneurs. Executives. Active LLM users since December 2022. Haven’t missed a single release. And they still went dark for a week.
The Real Breakthrough
The breakthrough isn’t Claude Code itself. It’s Skills.
A skill is a markdown file. You write instructions in plain text. Claude executes them. No Python. No APIs. No frameworks. Just instructions you’d give a smart colleague.
If you understand how a task gets done — you spend a few hours explaining it to the model. You get an agent that does that task for you. Again and again.
You don’t need to be a specialist. You need to know how to set the task and check the result.
Manager Becomes Executor
Used to manage programmers? Now you write code yourself.
Used to oversee lawyers on contracts? Now you review contracts yourself.
Used to brief designers? Now you make the mockups yourself.
The manager becomes the executor. (Whether that’s liberation or a different kind of trap — ask me in a year.)
One of my friends wrote: “A task that would have taken me a month — done in an hour.”
This feels like the biggest jump in generative AI since GPT-3.5.
Remember November 2022? GPT-3.5 dropped and Twitter was nothing but screenshots. The feeling that the world had shifted. Same thing is happening now. Just quieter.
People stopped keeping up. Models release faster than you can learn them. If you have any job besides tracking AI — you won’t keep up. (This includes me, and I’m supposed to be paying attention.) That’s not criticism. That’s diagnosis.
The Ceiling
Here’s what nobody tells you.
Skills work beautifully until they don’t.
Small skill — 50 lines — pure magic. You write, it works.
Bigger skill — 300 lines, five stages, parallel agents — black box. Something breaks. You don’t know where. You don’t know why. You change one line, wait 15 minutes, get wrong output, roll back, repeat.
Jake spent an hour debugging character hairstyles across slides in his visual novels. Which of the five stages broke? He had no idea.
Tyler’s skill deleted his humor reference examples at some point. Instructions still said “write black humor.” Examples were gone. He published 4-5 bad posts before noticing.
And then there’s the Salah problem.
My Wall
I was running 24 parallel agents analyzing data for a decision system.
One agent recommended a player who wasn’t even available. The model never received the “available options” list — so it invented a recommendation. Confidently. With stats.
Garbage in one agent propagated through all 24 dependent agents.
I spent hours trying to trace where the data broke. Which agent expected what format. Where the chain went wrong.
That’s when I knew: this can’t be fixed with more prompting. This needs visibility.
The Pattern
When I looked up from my screen, I saw my friends hitting the same wall.
Jake with his hairstyles. Tyler with his deleted examples. Different use cases. Same problem.
Then I started seeing posts from people I don’t know. Same enthusiasm. Same language. Same ceiling approaching.
The pattern was obvious. Skills cracked the creation problem. The maintenance problem was next.
Day One
Today is day one for Skillpad.
We built this for ourselves first. A visual workflow editor for Claude Code Skills. Open the markdown, see the graph. Stages, connections, data flow — explicit instead of implicit.
What it does today: visualization and editing. You can see your skill’s structure and modify it safely.
What it will do: regression detection, version comparison, execution tracing.
We’re three founders. Two technical, one (me) who kept drowning until we built a lifeboat.
If your skill is under 100 lines, you probably don’t need this yet.
If it’s over 300 lines and you’ve never debugged blindly for an hour — you’re either lying or lucky.
For everyone else: join the waitlist.
— Illia Martyn