The Claude Code Workflow Nobody Talks About: Building Skills From Your Own Sessions
Stop inventing skills. Your sessions already know what you need.
I sat down to write my first skill and stared at an empty Claude session.
Then I made one up, something that felt useful and looked great.
Claude ignored it completely.
The problem isn’t just the writing; it’s also where the intent came from. I invented it artificially, like most of us do when we make skills. The problem with these skills is they’re always generic because you’re guessing at what you’ll need instead of looking at what you actually do.
The good skills are already sitting in all of your sessions.
Your sessions already know where you struggle
Claude Code keeps track of how you work. Run /insights and it’ll give you a report. It’ll show you where you hit walls, correct things, and where you have to explain yourself all the time.
This report is your biggest insight into where you want to build your skills.
You don’t brainstorm skills. Instead, you harvest them.
Not everything should be a skill
This is the first mistake I made. I looked at my report and tried to turn everything into skills.
A lot of it is thinking about how you’re prompting, instead of adding skills.
Here’s a kind of pattern I use to understand whether it’s a skill, or rule, or a personal change:
Skill — a repeatable task you keep walking Claude through, step by step.
Rule — a one line correction you keep making. This goes in your CLAUDE.md, not a skill.
Personal — a note of these if they keep coming up. It’s something you have to change around your prompting and your workflow.
This is probably the first step in improving your workflow, and it’s basically sorting your feedback.
Where skills fall apart
You find a real skill, you’ll write it and save it.
And a lot of the time, Claude still won’t use it.
This is a very common complaint with skills.
Auto triggering only fires 30 to 50% of the time.
Personally, I always trigger skills myself rather than letting Claude try and pick it up.
Auto triggering sounds nice, but it’s not reliable enough.
If you do want it to auto-trigger
If you do want a skill to fire on its own and you want the highest chances of success, then there are two rules.
Descriptions will get cut off if they’re long, so make the important stuff come first.
Use extremely concise words that you would type. Instead of saying “execute this test suite,” just say “run tests.”
A lot of trigger failures actually do come down to vague and verbose descriptions that don’t sound like real user intent.
Keep it short, every line costs you
Everyone’s always heard that the CLAUDE.md file has to be very short.
The same goes for skills. It will sit in Claude’s context and get re-run every message.
A bloated skill isn’t very good. It will lead to more context rot.
So keep it lean. If it’s getting long, push the detail into a REFERENCE.md files and point at them (however, be very careful with using a lot of these files).
A skill isn’t done after one write
This is where you don’t get much help from any of the docs or anything like that.
Writing skills are the first step. Checking whether you use them is the second step.
So if you use it for a week, run insights again, and see if your friction changed based on the skill.
If it has, then it’s working. If it hasn’t, then the skill needs to be modified or you’re using it wrong.
Skills aren’t just to be left alone. You should constantly update them and improve them.
And they should be constantly reviewed to check if they’re doing something that’s worth it.
The models that we’re using have consistent updates, so the skills are going to be redundant when models perform better, and they’ll be needed when they perform worse.
If you take away one thing
You don’t write skills from imagination.
You watch what you keep doing, sort it honestly, and hand the repeatable parts to Claude in a way it’ll actually use.
And the quick test for whether something’s even a skill:
If you’d have done it from memory anyway, it wasn’t a skill. It was a note.
My skills
Everything I build is in my skills repo if you want to dig through it.
Star it if it’s useful, it tells me whether to keep building these.
Before you go
If you found this useful, please like and share, it means a lot to me and let’s me know I’m going in the right direction.
If you want more content, subscribe!
My inbox is always open, so come say hi.
Thanks for reading. I appreciate you.





