These past three weeks we’ve covered the early steps: leaving a chat for an agent, learning how to talk to it, and figuring out how it works.
That was all about what the agent can do. Now it’s time to cover something that actually changed the way I do my work. It’s called a skill. Honestly, it might be the most overpowered (yet simple) AI tool that you can master.
What is a skill?
Skill = instruction
Think of a skill as an instruction. A recipe.
Imagine you onboard someone new to help you. You don’t re-explain the job every time. The first time, you write a clean step-by-step guide: here’s the task, here’s where the files live, here’s what “done” looks like.
Next time you will just simply ask them to “do the same thing you helped me with last month”.
Your agent is that intern. A skill is that step-by-step guide.
Skill = prompt on steroids. And it pays off in one exact moment: when there’s a pile of context you’d happily explain once, for a task you do again and again. The skill captures it one time and reuses it forever.
The skill I use every week
It’s easier to show how cool it is on a real example.
Every week I pull numbers from 3-5 dashboards across business lines and turn them into one clean table with takeaways in bullet points. It took me about an 1-2 hours every other week.
Now it’s a skill. I type one command. A minute later the table is sitting there as well-formatted PPTX. Outliers/insights list is ready to review in text too.
Here’s what I packed into it:
The instruction - the steps: refresh the data, filter to my slice, clean it, build the table, write the takeaways.
A schema note - what each of the ~100 columns across dashboards mean, so the agent never guesses.
A context note - the abbreviations, the business “why,” what to do depending on the question I show up with that week.
The config - credentials to connect to dashboards directly on my behalf
The scripts - the exact queries and cleanup I used to run by hand, pre-written so the agent just runs them instead of reinventing them.
An example output - table example, formatting rules
I taught it once. Now I spend 1-2 minutes where I used to spend 1-2 hours.
The anatomy of a skill
Skill = SKILL.md + references + scripts + assets + configs
That example wasn’t special. Every skill has the same shape.
Skill is a folder
SKILL.md- the mandatory thin instruction up top. A name, one line on what it does, and the steps. It points to the other pieces instead of swallowing them. I keep mine under 300 lines (more on that below).references/- the deeper docs the agent opens only when it needs them. My schema note. My business context. The heavy stuff that shouldn’t clog the main file.scripts/- repetitive code the agent just runs. No reinventing a query it’s run fifty times.assets/- examples of the output, or templates for the input. The “here’s what good looks like.”<other>/- folder structure is flexible. You can have “configs” or “logs” or anything that is relevant to your skill.
I think of SKILL.md as the cover page and table of contents. Everything else is a chapter it opens only when the task calls for it.
How the agent uses them
In 2026 skills are becoming standard. Most agents read them natively.
Skill name/description is a “book cover” for your agent
A skill is invisible unless it sits in the right folder. No folder, no skill.
Those folders live at two levels.
User-level - in your home folder, every session sees them.
Project-level - tucked inside one project, visible only when you’re working there.
The paths look like
.claude/skills/<skill-name>or~/.agents/skills/<skill-name>or else depending on your agent.The agent reads a skill through gates, not all at once:
Gate 1 - at session start it only glances at each skill’s name and one-line description. Enough to know it exists and when it’s useful.
Gate 2 - the moment your ask matches one (or you type
/skill-name), it opensSKILL.md.Gate 3 - it reads only the references that task actually needs, and ignores the rest (or reads all if all’s relevant)
The agent skims the table of contents of every skill, and opens only the chapter the task needs.
That’s why SKILL.md has to stay thin. A bloated main file is like an intern who re-reads the entire handbook before telling you where the printer is (and the answer is wrong)
Why I thinks skills are overpowered
They fit you like nothing else. Everyone has unique workflows. A skill might be useless to everyone else and still save you an hour a week. That’s the point.
They tap your whole setup. My agent is already wired into my Gmail, calendar, database, and analytics. A skill can use it all of it at once.
They compound with routines. Pair a skill with a scheduled trigger and it runs the weekly job while I sleep.
They travel with you. When I moved between agents, my skills came along with minor folder (
.cursor/→.claude/) migration.They’re stupidly easy to share. Step 1 – send a skill zip/folder to your colleague. Done. Shared.
They turn slow work instant. When the input and output are standard/determined, a skill chains scripts and collapses hours into seconds - like printing a clean five-slide deck with fresh data in the time it takes to read this paragraph.
Skills are not perfect
It’s still a prompt, not a hard wall. I can write “never touch file X” in the skill - but if I then casually ask the agent to change it mid-conversation, there’s a real chance it listens to me over the rule. A skill steers behaviour; it doesn’t lock it.
Sharing on a team gets messy. I customise a skill, send it over, a teammate improves it, and now I want their version back… keeping everyone in sync takes real discipline.
There’s no universal playbook. Turning complex workflows into a skill can be tricky. Knowing whether a skill is actually good only came from building a few. Back in January I wrote 1,000-line
SKILL.mdfiles sitting alone in a folder, convinced they were brilliant. They weren’t. Thinking in small reusable pieces - references, scripts, assets that snap together - is a skill of its own (pun intended), and it only showed up with reps.
An Idea For This Week
If you take one thing from this issue, build a single skill. Shocking, I know. The simplest path I’d take:
Pick one workflow you repeat - something with tables, files, or reports is perfect.
Open Claude Code (or other agent) in the Desktop app (the easiest door, from last week).
Do it exactly as you normally would, but jointly with - tell the agent what you’re doing, where the file lives, what the output should be.
When you’re happy, ask it to package the flow into a skill. Most agents have a
/skill-creator(or/create-skill) command for exactly this.Done. Next week you summon it instead of rebuilding it.
The first time I built a skill it felt like overkill. The tenth time I used it, it felt like cheating.
My Shipyard Updates
Two things.
1/ Twitter Screenshot crossed 1,000 users. Boring SEO bet compounding. Watching so many people using it makes me happy
Can’t believe we grew from 100 in January to 1000 in May
2/ Learning product analytics for the first time. I embedded PostHog into a few products, so now I can see the user journey - what people click, where they linger, where they drop off. So much data – need to structure it well.
It’s not stalking if they accepted your cookies 😄
What’s the one workflow you’d never have to think about again if your agent just… knew it? Reply on this e-mail, I read everything.