# Where to find Claude Code skills in 2026: repos, directories and packs

> The four places Claude Code skills come from, what each one is good for, and the maintenance question that decides whether a skill still works next month.

- URL: https://www.agentskillpacks.com/blog/where-to-find-claude-code-skills
- Author: İsmail Günaydın
- Published: 2026-09-21
- Updated: 2026-09-21

## Quick answer

Claude Code skills come from four places: Anthropic's own repository, public GitHub repos from individual developers, aggregator directories that index those repos, and paid packs. All four install identically by dropping a folder into your skills directory. The difference is curation and whether anyone maintains them after the CLI changes.

Claude Code skills come from four places, and all four hand you the same thing: a folder with a `SKILL.md` inside it, which you drop into your skills directory. There is no format difference between a skill from Anthropic, a skill from a stranger's GitHub repo, and a skill someone charges for. Knowing that removes most of the confusion in this question before it starts.

What differs is curation and maintenance. Everything below is about those two.

## The four sources

**Anthropic's own repository.** [github.com/anthropics/skills](https://github.com/anthropics/skills) is the reference. Start here even if you install nothing from it, because reading a handful of well-formed skills teaches you the conventions faster than the docs do, particularly how the `description` field should be written. The [Agent Skills announcement](https://www.anthropic.com/news/skills) from October 16, 2025 covers the design intent.

**Individual developer repos.** The bulk of the ecosystem. Someone builds skills for their own work, publishes them, and you benefit. Quality ranges from excellent to abandoned, and the useful signal is not stars. It is whether the repo has commits after the last time Claude Code changed how skills behave.

**Aggregator directories.** Several sites index public skill repos and present them as browsable catalogues. Genuinely useful for discovery, especially for finding out that a skill for your niche already exists. Worth being precise about what they are, though: they index, they do not vet. A listing means a repo exists, not that it works or that anyone still looks after it.

**Paid packs.** Curated collections sold directly, including the ones I sell. Same files, same install, and you are paying for selection and upkeep rather than capability. I have written the honest version of [when that is worth money and when it is not](/blog/free-vs-paid-claude-code-skills), and the answer is more often "not yet" than a person selling packs would prefer.

## How installation works

The [skills documentation](https://code.claude.com/docs/en/skills) covers this, and it is short. Skills are discovered from two locations: your user-level skills directory, for things you want everywhere, and a project's `.claude/skills/`, for things that belong to one repository. Each skill is one folder containing a `SKILL.md`.

Two failure modes to know before you hit them, both of which I have caused:

- **A folder with no `SKILL.md` is invisible.** It sits there looking installed. When I audited my own install I found ten of these.
- **A `SKILL.md` with no `description` can never be selected.** I had six.

Neither produces an error message. That is worth internalising early, because it is the characteristic shape of every problem in this system: nothing breaks, things simply never happen.

## The question that actually predicts usefulness

Not stars, not skill count, not how good the README looks.

Ask: when was this last updated, relative to the last change in how Claude Code handles skills?

Skill discovery, frontmatter conventions and [settings](https://code.claude.com/docs/en/settings) have all moved since launch. A repo that predates a change you are hitting usually fails quietly rather than loudly, which means you can spend an hour concluding that skills are unreliable when the actual problem is that you installed something written against a previous version.

This is the single largest practical difference between the free and paid routes, and it is not about quality. A maintained repo and a maintained pack are equally fine. An abandoned repo and an abandoned pack are equally useless, and the pack cost money.

## Why "500 skills" directories are the wrong shelf to shop from

There is a pattern worth naming, because discovery sites are built to encourage it and the incentive runs against you.

Browsing a directory makes installing feel free. The skill is free, the install is one command, and the obvious move is to take anything that looks remotely useful. That intuition is wrong in a specific way: skills are selected by matching your request against every installed description, so each addition makes the selection problem harder for everything already there.

I measured my own install on September 21, 2026. Across 383 skills the descriptions alone came to 108,998 characters, and only 136 of them said when the skill should fire. Thirty-one names began with `blog`, twenty-six with `seo`, twenty with `geo`. Those clusters are not capabilities stacking up. They are near-duplicates competing, and the thirty-second blog skill makes the first thirty-one slightly harder to reach. The mechanism is in [why half your skills never fire](/blog/why-claude-code-skills-dont-trigger).

So the right way to use a directory is as a search engine, not as a shelf. Go in with a job. Take one skill. Leave.

## Reading a skill before you install it

Three minutes of reading beats any rating, and all three things are visible from the repo.

**The `description` field.** Open a `SKILL.md` and read it first. If it says what the skill is rather
than when it should fire, that skill will sit unused no matter how good its instructions are, because
selection happens on that field alone. This is the single highest-signal thing in the file and most
authors treat it as a subtitle.

**The length of the instructions.** A skill that is four hundred lines of prose is usually a document
someone converted rather than a workflow someone designed. The good ones tend to be short and
imperative, with the decisions made for you. Length is not quality here, and it is mildly
anti-correlated with it.

**Whether it says no anywhere.** Search the file for a refusal: a step that must happen before
another, a condition under which it stops. A skill with no refusal in it is a prompt with a folder
around it, and it will drift back into ordinary conversation the moment you push against it. That
drift is the most common disappointment with skills and it is visible before you install.

## What a directory listing does not tell you

Worth being explicit, because the interface implies otherwise.

A directory shows you that a repo exists and usually what its README claims. It does not tell you
whether the skill fires, whether it still works with the current CLI, or whether its description
competes with something you already have installed. None of those are things an index can know, and
all three decide whether the skill is useful to you specifically.

The last one is the one people never consider. A skill that is excellent in isolation can be useless
in your folder because a neighbour already claims the same requests. That makes "is this skill good"
the wrong question and "does this skill fit what I already have" the right one, and no directory can
answer the second.

## A sane process

1. **Read three skills in [Anthropic's repo](https://github.com/anthropics/skills)** before installing anything, to calibrate what good looks like.
2. **Write down the job** you want a skill for, in the words you would actually type.
3. **Search the directories** for that job. Ignore everything adjacent.
4. **Check the repo's last commit** before installing.
5. **Install at most three, use them for a week**, then decide what stays.
6. **Delete aggressively.** An unused skill is not free, because it competes in every selection.

Step six is the one nobody does and the one that keeps the folder working.

If you would rather skip the curation entirely, that is what a pack is for: [CodeForge](/products/codeforge) for the core development workflow, [AI SEO Command Suite](/products/ai-seo-command-suite) if SEO operations are the recurring job. The [six I actually use](/blog/best-claude-code-skills-for-indie-developers) has the full list, with the conflict of interest stated where you can see it.

## FAQ

### Where do Claude Code skills come from?

Four sources. Anthropic's own public repository, individual developers publishing repos on GitHub, aggregator directories that index those repos, and paid packs sold directly. All of them produce the same thing, a folder containing a SKILL.md file, and all install the same way.

### Is there an official Claude Code skills marketplace?

There is an official repository from Anthropic, which is the reference for the format. The sites describing themselves as marketplaces are mostly indexes over public GitHub repositories, which is useful for discovery and is not the same as vetting.

### How do I install a skill I found?

Copy the skill's folder into your skills directory, either the user-level one for everything you do or the project's own for a single repository. The CLI picks it up from there. Restart if it does not appear.

### How do I tell whether a skill is maintained?

Check the last commit date against the most recent Claude Code release that changed anything about skills. A repo that predates a format change will usually fail silently rather than error, which is the worst kind of failure to debug.

### How many should I install at once?

Three, then use them for a week. Skills compete for selection, so a bulk install degrades the behaviour of everything already in the folder. The instinct to grab fifty at once is the single most common mistake here.