My 5 lessons from using AI skills

I’ve spent enough time using AI tools—Claude, ChatGPT, and other models—to get past the initial “this is amazing” phase.

At first, I mostly treated AI as a smarter search box. I asked questions, generated drafts, summarized things, and occasionally used it to help with code.

Then I started experimenting with AI skills: giving models access to tools, files, structured workflows, and repeatable instructions so they could actually do things rather than just tell me how to do them.

That changed my view of AI quite a bit.

Here are five lessons I’ve learned.

1. The model matters less than the system around it

I used to spend a lot of time comparing models.

Which one writes better? Which one codes better? Is Claude better at this? Is another model better at that?

Those differences are real, but once I started building repeatable skills and workflows, I realized something: the model is only one part of the system.

A great model with poor context, vague instructions, and no access to the right tools can produce surprisingly mediocre work.

A slightly less capable model with the right files, clear instructions, good examples, and a well-designed workflow can be incredibly effective.

So I’ve started thinking less about “Which model is smartest?” and more about “What does the model need in order to succeed at this task?”

That has been a much more useful question.

2. The best skills are single-responsibility tasks with clear input/output schemas

My first instinct was to build ambitious skills.

“Research this market.”

“Help me run this project.”

“Analyze my business.”

They sounded impressive, but they were too broad. There were too many decisions hidden inside each instruction.

The skills that actually became useful were much narrower.

Take a meeting transcript and turn it into decisions, owners, and next steps.

Read a set of customer comments and group recurring complaints.

Review a document using the same checklist every time.

Take a rough brief and turn it into the structure my colleagues normally use.

None of these sound revolutionary.

That’s exactly why they work.

I’ve learned that if I can describe the input, the process, and the expected output clearly, I can usually build something reliable. If I can’t explain those things, I’m probably asking the AI to make too many judgment calls on my behalf.

3. Giving AI more autonomy doesn’t automatically make it more useful

There’s a natural temptation with AI agents and skills: give them more access.

More tools. More permissions. More steps they can execute independently.

I assumed that would always make them more capable.

It doesn’t.

Every additional tool also introduces another opportunity for the model to misunderstand what I wanted, choose the wrong action, or confidently continue down the wrong path.

I’ve had workflows where the AI successfully completed six steps—only for me to realize that step two was based on a bad assumption.

Now I prefer constrained autonomy.

I want AI to move quickly through reversible, low-risk work. I want checkpoints around decisions that are expensive, external, or difficult to undo.

The goal isn’t maximum autonomy.

It’s useful autonomy.

That distinction has saved me a lot of frustration.

4. Context is probably the best thing I can give an AI

I used to blame the model when an output wasn’t good.

Sometimes that was justified. Often, though, I simply hadn’t given it enough information.

“Write this in our style.”

What style?

“Analyze these results.”

Against what objective?

“Make this presentation better.”

Better for whom?

Humans fill in enormous amounts of context without noticing. AI systems often need that context made explicit.

Once I started giving models examples of previous work, definitions, constraints, audience information, templates, and explanations of what “good” looks like, the quality improved dramatically.

This has changed the way I think about prompting.

I don’t spend much time searching for magical prompt formulas anymore.

I spend that time improving the context.

5. AI is most valuable when it becomes part of a workflow I actually repeat

I’ve built plenty of clever AI experiments that I used exactly once.

They were technically impressive and practically irrelevant.

The useful ones are much less exciting.

They save me 15 minutes every Monday.

They remove a tedious step from something I do every week.

They make the first draft of a recurring task consistently better.

They help me process information that I otherwise would have postponed.

That’s where the value starts compounding.

If a skill saves me 20 minutes once, that’s nice.

If it saves me 20 minutes three times a week—and produces a more consistent result—that starts changing how I work.

So now, when I experiment with AI, I ask myself a simple question:

Will I actually use this again?

If the answer is no, it might still be an interesting demo.

But it probably isn’t a useful skill.

What I’ve changed my mind about

After using AI this way, I’m less interested in AI that can do everything.

I’m more interested in AI that can do a few things reliably inside the way I already work.

The biggest gains haven’t come from writing longer prompts or constantly switching to the newest model.

They’ve come from understanding the task better.

What information does it need?

What decisions should the AI make?

What decisions should stay with me?

What does a good result actually look like?

And where does this fit into a workflow that happens repeatedly?

That’s the less glamorous side of using AI skills.

It’s also where I’ve found most of the real value.