Free vs paid Claude Code skills: what the free ones actually cost

Most Claude Code skills are free. You clone a public repository, drop the folders into ~/.claude/skills/, and the CLI picks them up. No payment, no account, no licence. For a large amount of work that is the correct answer, and I sell skill packs, so take that sentence as the one I had the most incentive not to write.

What the free route costs is not money. It is a resource that is harder to see on an invoice, which is why the conversation about it barely exists.

What a skill is, mechanically

A skill is a folder containing a SKILL.md file: YAML frontmatter with a name and a description, then Markdown instructions underneath. Anthropic introduced Agent Skills on October 16, 2025, and the documentation covers the layout. Skills live either in ~/.claude/skills/ for everything you do or in a project's .claude/skills/ for one repository.

There is no paid tier in the format. A skill someone charges $19 for and a skill from the official Anthropic repo are the same kind of file, loaded the same way. Anyone implying otherwise is selling you a distinction that does not exist.

Which means the honest comparison is not free versus paid capability. It is about what happens when you have a lot of skills, and here the free route has a specific failure mode.

What 383 installed skills actually cost

I measured my own install on September 21, 2026 rather than reasoning about it, because I had been assuming the cost was small and wanted to know.

Read the fourth row first. Roughly 109,000 characters of description text exists to answer one question: which skill applies to what you just asked for. At the usual English ratio that is somewhere near 27,000 tokens of routing overhead, spent before any actual work begins.

Then read the last row. Only 136 of 383 descriptions say anything about when the skill should be used. The remaining 247 describe what the skill is, which is a different thing and a much less useful one, because matching a request to a capability is exactly the job the description exists to make easy.

And the first two rows are the free ecosystem in miniature: ten directories that cannot load at all, and six skills with nothing to match against. They were installed, counted, and inert.

The cost is a choice problem, not a storage problem

Here is the part I had wrong, and I think most people have wrong.

I had been thinking about a large skill install the way you think about installed apps: they sit there, they cost disk, and they are free until used. That intuition is wrong in a specific way. A skill is not dormant. It participates in every decision about what to do next, because that decision is made by reading descriptions and picking one.

So the real cost curve is not linear in install size. It is worse than linear, because the failure mode is not "slower", it is "chooses wrong". Thirty-one of my installed skills start with blog, twenty-six with seo, twenty with geo. Those clusters overlap heavily in what they claim to do. Adding the thirty-second blog skill does not add a capability so much as it adds a near-duplicate for the model to distinguish from thirty-one others, using descriptions that mostly do not say when to fire.

That is why "500 skills in one bundle" is a warning label rather than a feature, and it is the one genuinely important thing to understand before you go shopping in either direction.

So when is free the right answer?

Most of the time, and specifically in these cases:

• You know exactly what you want the skill to do. If you can describe the workflow, a free skill that implements it is as good as a paid one.
• You are exploring. Nobody should pay to find out whether skills fit how they work. Install a handful of free ones and use them for a week.
• Your need is narrow. A single well-written skill for one job beats a pack you bought for one of its forty skills.
• You are willing to maintain them. Free skills are yours to fix when the CLI changes. That is a fair trade if you enjoy it.

I would go further. If you are researching this question at all, you are probably at the stage where the answer is free, because the thing a pack sells is a shortcut past decisions you have not yet made badly enough to want removed.

So what is worth paying for?

Three things, and capability is not among them.

Curation. The value of a good pack is the skills that are not in it. Twenty skills chosen to cover a workflow without overlapping beat two hundred that collectively cover more and individually confuse the router. You are paying someone to have deleted things.

Maintenance. Claude Code changes. Skill discovery, settings and frontmatter conventions have all moved since launch, and a repo whose last commit predates the change you are hitting will fail quietly rather than loudly. A maintained pack is someone else's job to keep current, which is worth actual money if your alternative is debugging a silent no-op at 11pm.

Enforcement. This is the one that justifies the price when it is real. Most skills describe a workflow. A few refuse to continue when you skip a step in it, and that refusal is the entire product, because the steps a solo developer skips are exactly the ones nobody is there to insist on. CodeForge is the pack I built around that idea: planning gate before the feature, test before the implementation, review after the edit, and it stops rather than drifting back into chat when you push against it.

The same logic applies per domain. AI SEO Command Suite is worth its price to someone running SEO operations weekly and worth nothing to someone who needs one audit. Humanizer Pro is three skills, and three is the point.

The question that settles it

Skip the comparison table. Ask one thing instead: what step do I keep skipping?

If you can name it, a pack that enforces that step is worth money, and you will know within a week whether it does. If you cannot name it, no pack will help, because you are shopping for capability and capability is free. You will install it, admire it, and never fire it, joining the 247 descriptions in my own install that never told the model when they applied.

That is also the honest reason I put this post in front of the six packs I actually recommend rather than after them. If the free route works for you, take it. The packs are still there when it stops.

Related reading: why half your installed skills never fire goes into the description-writing problem, and where to find Claude Code skills covers the repos and directories worth knowing about.