How to edit a Claude Skill

Most people build a Claude skill, get it working, and never touch it again. That's understandable. You spent time setting it up, it's producing results, and there are plenty of other things demanding your attention.

But a skill that was right for you six months ago might not be right for you now. Your workflow changes. The way you communicate shifts. What you need from a skill evolves, and the instructions don't update themselves.

Why Skills Go Stale

The problem is subtle. The skill still works. It's just not working as well as it could.

You might notice the output feels slightly off, or you keep making the same manual edits before you publish or send. Maybe it's taking you longer to get from the skill's output to something you're actually happy with. These are signs the underlying instructions need attention, not that the skill is broken, just that it's drifted from what you currently need.

It's worth thinking about a skill the way you'd think about any working process. You'd revisit a process if it stopped delivering the results you wanted. Skills are no different, and the consequences of leaving them unchanged are the same: a gradual drop in quality that you compensate for manually, without ever fixing the root cause.

Ask Claude to Review the Skill

Here's what I do when I suspect a skill needs work. I open a new chat, give Claude the skill name, and ask it to review it.

The prompt is simple: "Review this skill. Where is it strong? Where is it falling down? What changes would make it better?"

Claude reads the actual skill file, the instructions, the tone guidance, and the output format, and comes back with specific suggestions. Not vague feedback. Actionable points tied to how the skill is currently written.

When I did this with my Italian teacher skill recently, it identified three changes that would meaningfully improve how the lessons were structured. I hadn't spotted them myself, partly because I was too used to how the skill already behaved.

When something mostly works, you stop questioning it. Having Claude look at it fresh gives you the kind of outside perspective that's hard to generate yourself.

How to Make the Changes

Once you have Claude's suggestions, you decide which ones to act on. Not every suggestion will fit; you know your workflow better than Claude does. But there are usually one or two things that are immediately worth fixing.

From there, it's straightforward. Ask Claude to implement the changes. It reads the current skill file, makes the edits, and packages up the updated version. You review it, click save, and that's it. Next time you run the skill, the new instructions are in play.

No separate tool. No technical knowledge required. The whole process typically takes less than ten minutes.

What to Check When You Review a Skill

A few questions worth working through when you go back into a skill:

Does the output still match how you want to communicate? Tone evolves. If your brand positioning has shifted since you first built the skill, the instructions probably haven't caught up.

Are you regularly editing the output before you use it? If you're making the same manual tweaks every time, those tweaks belong in the skill instructions, not your workflow.

Has your process changed in a way that affects what the skill needs to produce? A skill built for one version of a workflow won't automatically adapt when the workflow changes.

Has your audience or platform changed? A skill written for one context, a certain channel, a certain type of reader may need updating if that context has shifted.

These aren't complicated questions. Running through them every few months is enough to keep a skill doing what you actually need it to do.

Don't Set and Forget

The skills you use most are the ones most worth maintaining. Go back into them occasionally. Ask Claude to review them. Make small adjustments before the gap between what the skill does and what you actually need gets wide enough to matter.

If you've built skills and haven't looked at them since, pick the one you use most and start there. Ask Claude to review it. See what comes back. You'll probably find at least one or two things worth updating, and you'll have a better-performing tool by the end of the day.

That's a good use of ten minutes.

Check out this Digital Spark where I show you how to build your first skill.

FAQ's

Q1: How do I know when a Claude skill needs updating?

A: The signs are subtle. Output that feels slightly off, making the same manual edits before you use it, or taking longer to get from the skill's output to something you're happy with. These mean the instructions have drifted from what you currently need, not that the skill is broken, just that it hasn't kept pace with how you work.

Q2: Can Claude review one of its own skills and suggest improvements?

A: Yes. Open a new chat, give Claude the skill name, and ask it to review it. A prompt like "Where is it strong? Where is it falling down?" works well. Claude reads the actual skill file and returns specific, actionable suggestions tied to how the instructions are currently written not vague feedback.

Q3: How long does it take to update a Claude skill?

A: Usually less than ten minutes. Once you have Claude's suggestions, decide which to act on, then ask Claude to implement the changes. It reads the current file, makes the edits, and packages up the updated version. You review it, save it, and the new instructions are live the next time you run the skill.

Q4: What should I check when reviewing a Claude skill?

A: Four things: whether the output still matches how you communicate, whether you're regularly editing outputs before use (those edits belong in the instructions), whether your workflow has changed in a way that affects what the skill needs to produce, and whether your audience or platform has shifted since you first built it.

Q5: How often should I review my Claude skills?

A: Every few months is enough for most skills. Prioritise the ones you use most; they have the highest impact on your daily output, and any drift compounds faster. If you're making the same manual tweaks to outputs regularly, that's a clear signal to check the skill sooner rather than later.