Claude Skills Clean-Up: Which Skills to Keep, Merge or Delete

Key takeaways

• Anthropic donated Agent Skills to the Agentic AI Foundation on 22 October 2026, the same home it gave MCP
• Every installed skill adds its name and description to each session, so a big pile adds up
• Keep skills that run exact code, hold private know-how or navigate an API with no connector
• Ask three questions: exact code? private knowledge? already replaced by a connector?
• Many habit-style skills, such as “grill me”, can become one line in AGENTS.md
• Mark’s own audit of 374 skills flagged 70 to delete and 46 to merge into 6
• Store seasonal skills in a separate folder and load them just in time
• Always run the audit read-only and review before deleting

If you have been using Claude Code or Codex for a while, your skills folder has probably grown. A new skill trends, you install it, use it a few times and forget about it. Months later, it is still there.

This article walks through a practical Claude skills clean-up, based on Mark Kashef’s new video and the official Agent Skills announcement. You will learn which skills to keep, which to merge or delete, and how to run a safe, read-only audit of your own setup.

What did Anthropic actually donate?

Anthropic handed the Agent Skills standard to the Agentic AI Foundation (AAIF), a Linux Foundation project, on 22 October 2026. That is confirmed in Anthropic’s announcement and the Linux Foundation press release.

A skill is a folder with a SKILL.md file that teaches an agent a task, plus any scripts or reference files it needs. Anthropic says more than 70 agent platforms now support the format, including ChatGPT, Cursor, GitHub Copilot and VS Code.

The donation covers the name, logo, GitHub organisation and the agentskills.io domain. You can keep using skills exactly as before. Nothing in your setup changes overnight.

Why does a big skills folder hold your AI back?

Every skill you install adds a little to every session, and lots of overlapping skills give the model more chances to pick the wrong one. Under the Agent Skills format, an agent first reads each skill’s name and description, then opens the full file when a task seems to match.

One skill costs very little. A hundred skills means a hundred descriptions competing for attention. Mark puts it simply: with 100 skills, you have 100 chances to pick the wrong one.

Overlap is the bigger problem. Mark describes having six design skills that all said slightly different things. Some sessions picked up the good one. Others picked up an older one, and the results got worse, not better.

Plugins add to the pile too. In Mark’s audit, about half of the skills on that machine came from plugins rather than anything he had installed on purpose.

Which three types of Claude skills are still worth keeping?

Keep skills that run fixed code, hold knowledge the model cannot find, or guide it through an API that has no connector. Mark groups the skills that still earn their place into these three types.

Even the second type can shrink over time. Mark’s view is that as built-in memory in AI tools improves, many personal skills could become a line or two in a memory or instruction file.

What three questions should you ask about every skill?

Ask: does it run exact code, does it know something the model cannot find, and has a connector already replaced it? If the answer to all three is no, the skill is a candidate to delete or shrink.

1. Does it run exact code? If it relies on scripts that must behave the same way every time, keep it.
2. Does it hold knowledge the model cannot get? Private processes, client preferences and house style count. General best practice does not.
3. Is there now a connector or MCP server for this? If the service now has an official connector, the skill may be out of date.

These questions are about skills you already have. If you are deciding whether to install something new, our 3-question test for new AI skills covers that side.

When can a skill become one line in AGENTS.md?

A skill can become one line when its value is a behaviour, not a body of knowledge or code. Mark’s example is the popular “grill me” skill, created by Matt Pocock, which asks the AI to quiz you before it starts building.

Mark says the current version is about 330 words. He replaced it with a single instruction in AGENTS.md, roughly: when my instructions for a new build are unclear, ask me a series of questions first.

In his side-by-side test, the questions from both versions had similar depth. The main differences were formatting, such as emojis and question rounds, which he says you can add with a few more words.

We have written a step-by-step method for this, including how to test the result fairly: How to Replace an AI Skill With One Line in AGENTS.md.

How do you merge overlapping skills?

Group skills that do the same job, keep the best instructions from each, and combine them into one compound skill. The aim is one clear source of truth per job.

1. List skills by job, such as design, writing or research.
2. Pick the strongest one in each group as the base.
3. Copy across only the unique rules from the others.
4. Remove contradictions, so the model never gets two answers to the same question.
5. Test the merged skill on a real task before removing the originals.

What did the 374-skill audit find?

On one of his Mac minis, Mark’s audit found 374 skills, recommended deleting 70, and suggested merging 46 into 6 compound skills. It was unsure about 25 more and left those for him to decide.

The report also showed his top 10 most-used skills and a heat map of when each skill was used. That makes the 80/20 pattern easy to see: a small group of skills does most of the work.

Mark points out this was a machine with old experiments on it, not his main computer. So treat these numbers as one person’s example, not a typical result.

How do you run a read-only skills audit?

Ask your AI to review every skill it can load, check your past sessions for usage, and produce a single report, without changing anything. Mark offers his full prompt as a free download from his video description. Here is a short version in our own words:

“Audit every skill my AI tools can load. Use my past sessions to work out which ones I actually use. Group overlapping skills, sort each one with three questions (exact code, private knowledge, replaced by a connector) and suggest keep, merge or delete. This is read-only: do not delete, move or edit anything. Give me one self-contained HTML report I can open in my browser.”

If your setup includes rules, memory files and plugins as well as skills, our guide to auditing your Claude and Codex setup covers the whole picture.

How do you keep occasional skills out of the way?

Move them into a separate folder that your AI tool does not load at start-up, then ask it to read a skill from there when you need it. Mark calls these just-in-time skills.

This suits skills you use in busy seasons, such as quarterly reports or launch weeks. They stay safe and ready, but they are not sitting in every session’s context.

The full set-up is in our guide: How to Load AI Skills Only When You Need Them, Not Every Session.

What does the video not confirm?

• The audit numbers (374, 70, 46 into 6) come from one of Mark’s machines and have not been independently checked.
• Figures about the “grill me” skill, such as its word count and GitHub popularity, are as stated in the video.
• His one-line test is a single comparison, not a benchmark. Your results may differ by model and task.

How should you start your Claude skills clean-up?

Start with a read-only audit, then make small, reversible changes. Here is a routine you can finish in an afternoon:

1. Back up your skills folders so you can undo anything.
2. Run a read-only audit with a prompt like the one above.
3. Delete the obvious dead weight: skills you installed to try once.
4. Merge overlapping skills into one per job.
5. Turn habit-style skills into one line in AGENTS.md and test them.
6. Move seasonal skills into a library folder for just-in-time use.
7. Repeat every few months, or when you add a batch of new plugins.

Who is this clean-up right for?

This clean-up suits you if:

• You use Claude Code, Codex or another agent tool most days
• You have installed skills or plugins from YouTube, GitHub or newsletters
• Results feel inconsistent from one session to the next
• You run several businesses or clients from one setup

It matters less if you only use a chat app with no skills installed, or you have a handful of skills you wrote yourself and use every week.

Frequently asked questions

The bottom line

Agent Skills moving to the Agentic AI Foundation does not break anything, but it is a sensible prompt to tidy up. Keep the skills that run code, hold private know-how or navigate an API, and question everything else. Run a read-only audit, merge overlaps, turn habits into single lines, and load seasonal skills only when you need them. Your first step: back up your skills folder and run the audit today.

Related posts

• How to Audit Your Claude and Codex Setup and Clear Out AI Clutter
• Should You Install a New AI Skill? A Simple 3-Question Test
• Claude Duet: Anthropic’s Proactive AI Agent, Plus Gemini 4 Carbon’s New Checkpoint

Garry Knight

I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

See Full Bio