How to Install a Claude Skill (and When It Is Time to Remove It)

You install a skill by creating a folder containing a SKILL.md file in one of three places: ~/.claude/skills/ for your own, .claude/skills/ inside the repository for project ones, or by installing a plugin from a marketplace. Claude finds it on its own, you do not have to register it anywhere. From that moment it costs you something in every session, even if you never use it.

That last sentence is why I wrote this article. The hard part is not installing a skill: it is deciding which ones to stop keeping.

How do you install a skill on Claude?

There are three routes, and they change who sees it.

Personal skill. You create ~/.claude/skills/skill-name/SKILL.md. Only you see it, on that machine, in every project. It is the right place for things that concern how you work, not how somebody else works.

Project skill. You create .claude/skills/skill-name/SKILL.md inside the repository and commit it. Anyone who clones the project gets the same one. It is the right place for the conventions of that codebase: how a deploy is done, how a commit message is written, how a new page is added. This site's repository has eight of them, and they exist so that a newcomer does not have to read thirty pages of documentation.

Plugin from a marketplace. You add a marketplace and install a package that brings several skills in at once.

claude plugin marketplace add anthropics/claude-plugins-official
claude plugin install plugin-name@claude-plugins-official
claude plugin list

You do not declare the skill in any configuration file. Claude reads the name and description of every skill it finds, and decides on its own when to open one. If you want to write your own instead of installing it, the full procedure is in the guide to building a skill from scratch.

What a skill actually is, and why it costs you even when unused

Put simply, a skill is nothing more than a large, well-structured, modular prompt, which can also contain software and connectors the agent calls. It is not magic and it is not a plugin in the old browser-extension sense: it is written instruction, organised so the model opens it only when needed.

Here is the part almost nobody explains. The body of the skill loads on demand, but the name and the description sit in context permanently, because that is how the model knows the skill exists. A skill you never use does not cost zero: it costs its own announcement, in every single session, forever.

The Anthropic documentation puts a number on it and calls the mechanism progressive disclosure: the first level, name and description, always loads at startup and costs around 100 tokens per skill; the body only enters context when the skill fires. That same page draws the optimistic conclusion, that you can install many skills without a context penalty. Which is true, one skill at a time.

You can do the maths with a command Claude Code already ships.

claude plugin details plugin-name@marketplace

On my machine, on 17 September 2026, the installed plugins alone declare 17,809 tokens permanently in context for 65 skills, roughly 274 tokens per skill, before I have typed a single line. On top of that sit the 63 personal skills in ~/.claude/skills. That is a hundred and twenty-eight skills the model carries at every start.

Those 274 tokens per skill are nearly triple the 100 quoted as the reference figure, because marketplace skills carry long descriptions: the priciest one in my set spends 1,165 characters on description alone, and I have never opened it. The per-skill penalty stays small, exactly as Anthropic says. It is the total that stops being small.

What happens when you install too many?

What I observed, before going to measure it: the agent gets tired. It reasons more slowly, it fires skills more slowly, and every so often it picks the wrong one, or fails to fire the right one at all. This is not an isolated impression, and it is documented.

The two studies that follow measure tools, not skills. The mechanism is the same though: a list of options, each with its own description, that the model has to choose from on every turn. And MCP tools are the closest analogue to a skill that research has actually studied.

The work closest to this case is RAG-MCP by Tiantian Gan and Qiyao Sun (arXiv 2505.03275, May 2025), which names the problem outright: prompt bloat. As the number of available tools grows, the model struggles to use them. By retrieving only the relevant tools first and showing the model just those, selection accuracy goes from 13.62% to 43.13%, more than triple, and prompt tokens drop by over 50%.

A second study, by four researchers at Meta Platforms (arXiv 2605.24660, May 2026), measures the same thing from another angle: on a catalogue of 370 tools, an agent presenting an average of 7.4 reaches 90.3% coverage, practically identical to the 90.8% of a fixed 50-tool list. Seven well-chosen tools are worth fifty sitting there.

Underneath sits the phenomenon you already know if you have read the eight skills from my workflow: the context rot measured by Chroma Research in July 2025 across 18 frontier models, where every single model degrades as context grows. A badly built skill does not help the model, it distracts it.

A number of my own, which I am stating for what it is: an estimate based on how I work every day, not a controlled experiment. Keeping the skill set pruned and coherent is worth me between 20% and 30% more available context compared to letting it grow unchecked.

How do you decide what to keep and what to remove?

I use three criteria, in this order.

Actual usage. Not the usage you imagined when you installed it. If you have not opened it in a month, its announcement is taking up context for nothing.

The risk to benefit ratio. The risk here has a precise name: the degradation you get once context fills up with pieces that sit there and bring no benefit. A skill that saves you ten minutes once a quarter, while making the model less precise in its choices every day, is a net loss.

Scope coherence. This is the criterion I use most and the one I see least often. An agent resembles a personal assistant more than a piece of software: you give it the competences of its job, not every competence that exists in the company. The agent that works on administration gets the administration skills, not the marketing ones and not the development ones. You do not install everything on everyone. This keeps each agent inside a narrow vocabulary, and the context efficiency comes from there before it comes from the raw count.

The technical counterpart of this criterion is exactly the Meta result (Repantis et al., arXiv 2605.24660): the selection should happen upstream, not be left to the model at the moment it has to choose.

When does a skill stop being useful?

Back in March 2026 I had listed frontend-design among the ones to use with care. Six months later I uninstalled it completely, for two reasons.

The first is that models improved on exactly that ground. Interface design is one of the things agents now do well on their own: the skill was explaining something the model had already learned.

The second is the bill. That skill carried a lot of data, and claude plugin details says it without argument: 78 tokens permanently in context, and roughly 2,700 tokens every time it fires. Two thousand seven hundred tokens for a contribution that had become marginal, while the model slowed down and worked less token-efficiently.

That is the general lesson here. A skill is not wrong forever or right forever: the model changes, and what filled a gap six months ago now fills it twice.

Frequently asked questions

How do you install a skill on Claude Code?

You create a folder containing a SKILL.md file in ~/.claude/skills/ if you want it just for yourself, or in .claude/skills/ inside the repository if it should apply to the whole project and to anyone who clones it. You do not need to register it in any configuration file: Claude Code finds it on its own at the next start. The third route is installing a plugin from a marketplace with claude plugin install, which brings several skills in at once.

Where do you find ready-made Claude skills?

In plugin marketplaces. The official Anthropic one is added with claude plugin marketplace add anthropics/claude-plugins-official, and independent ones exist, reached through their Git repository address. claude plugin marketplace list shows the ones you have configured and claude plugin list the ones installed. Before installing a whole package, check how many skills it brings: they all come in, and they all take up context.

How many skills can you install on Claude?

There is no technical limit, and that is precisely the problem. The limit is practical: the more options the model has to choose from, the worse it chooses. The RAG-MCP research shows selection accuracy going from 13.62% to 43.13% when the model is shown only the relevant tools, and the Meta work reaches the coverage of a 50-tool list with 7.4 tools. Count what they cost you with claude plugin details, not how many you have.