Free vs Paid Claude Code Skills: When $9 Beats Free

Free Claude Code skills are the right requisition more often than a storefront wants to admit. If a SKILL.md is short enough to read top to bottom and the author is someone you trust, take the free one — paying buys you nothing. Paid earns its keep on a different axis: verification at scale, maintenance, and proof a skill works before you run it. A skill executes with your agent's permissions, and one public audit of a scraped free set counted 140,963 issues across 22,511 skills, roughly 36% carrying prompt-injection patterns. When the cost of vetting an unverified file exceeds the price of a tested one, $9 beats free. Here's the honest breakdown, sorted by which side actually wins.

When free Claude Code skills are the right call

Free wins outright in these situations. Don't let anyone upsell you past them.

• You can read the whole thing. A 40-line SKILL.md you scan top to bottom carries no hidden instructions — you've verified it yourself. That's the strongest safety check there is. Requisition it and move on.
• You trust the author. A maintainer with a track record, or a file you've already run before, needs no purchase to justify it.
• It's a one-off. Throwaway task, no reuse, no maintenance burden. Grab a free file, use it once, delete it.
• Anthropic ships an official one. The anthropics/skills repo is free, first-party, and compatible with Claude Code, claude.ai & Codex. When it covers your job, start there — a depot has no business competing with the first-party source.

If your situation is on that list, close the tab. The rest of this guide is for the situations where it isn't.

When paid Claude Code skills earn their keep

Paid is not "better skills." It's a different guarantee. Four things you're actually buying:

Verification at scale. The free directories' strength is also their weakness: volume without a check. The largest holds on the order of 2.2 million SKILL.md files. Volume is the metric those catalogs optimize — it's also the metric that hides the bad file. The audit numbers above aren't a scare tactic; a skill runs with your agent's permissions, so a poisoned file isn't a cosmetic defect. It's the difference between reading a manual and letting a stranger operate the equipment.

Maintenance. A free file is frozen at whatever commit you grabbed. When the runtime shifts or an App Store guideline changes, nobody patches it. A maintained skill gets updated, and a locker re-download keeps you on the latest version instead of a copy quietly rotting in your repo.

Proof before you run it. Most catalogs show a description that claims the skill works. AgentSource attaches a field report to every skill — real captured output from a documented run, readable before you pay. The Deep QA Audit Loop report, for example, shows a live finding where a nutrition app read "goal met" for brand-new users across 13 call sites. That's proof no star rating and no bulk listing can fake.

Time. Vetting an unverified skill line by line — especially one with write, delete, or network permissions — takes longer than $9 of anyone's hour. If you can't fully audit it and can't afford to run it blind, a tested one is the cheaper option, not the pricier one. The MCP Tool Reliability Auditor exists precisely because tool definitions that look fine hide auto-fail defects until something deletes a record.

Free vs paid: the honest landscape

Every option below is legitimate. They optimize for different things, and knowing which is the whole game.

Option	What it is	Price	Genuinely wins at	Watch for
Bulk free directories (SkillsMP ~2.2M, LobeHub ~169k, claudemarketplaces.com ~23k)	Scraped / aggregated SKILL.md files	Free	Selection, zero cost, fast browsing	No verification at scale — one scraped-set audit found 140,963 issues across 22,511 skills, ~36% with prompt-injection patterns
Multi-vendor marketplaces (Agensi ~70% payout, ClaudeSkills.ai ~90%)	Many creators, one storefront	Varies per seller	Variety, an active creator ecosystem	A headline payout recruits vendors, not buyers — quality varies seller to seller
Single-seller bundle (e.g. a ~$99 one-time skill stack)	One author's curated set on their own site	~$99 one-time	Proves the paid model works; one coherent voice	One sales page — no per-skill page, no pre-purchase proof per skill
Official repo (anthropics/skills)	Anthropic's own first-party skills	Free	First-party, trustworthy, compatible everywhere	Scope is what Anthropic chose to ship — not your specific niche job
AgentSource depot	14 curated skills, one factory	$9 / $29 / $49, one-time	A field report per skill (read before buying), maintained, one indexable page each	Small catalog by design — not the place for breadth

The free directories genuinely beat a paid depot on selection and cost. A multi-vendor marketplace beats it on variety. A single $99 bundle beats it on all-at-once value if you already trust the author sight unseen. None of that is in dispute. What a depot trades on is the one thing the others structurally can't: a readable proof-of-work attached to each item.

Decision table: free or paid?

Your situation	The call
SKILL.md is short enough to read end to end	Free — you've verified it yourself
Author is first-party or someone you already trust	Free
One-off task, no reuse, no maintenance	Free
Anthropic ships an official skill for the job	Free — start at anthropics/skills
Skill has write/delete/network permissions you can't fully audit	Paid or independently vetted
You'll rerun it across many projects over months	Paid — maintenance and updates matter
You need proof it works before you run it	Paid — look for a field report
Your time to vet it exceeds the price of a tested one	Paid — $9 beats the hour

Where AgentSource fits

Not free-bulk, not a multi-vendor bazaar, not a one-page bundle. AgentSource is a single-source depot: a small rack of skills, each one curated, tested, documented with a field report, and maintained by the same factory that shipped them in anger. Pricing is one-time — $9 a single, $29 a category stack, $49 for the Full Kit — with a 7-day no-questions refund and a locker that re-issues the latest version whenever you need it.

That's the honest position: for breadth and zero cost, the free directories win, and you should use them for anything you can read yourself. When you need to know a skill does the job before you hand it your agent's permissions, that's what the field report is for. Two axes, two tools. Pick by the job in front of you — and if you're still calibrating what a skill even is, the SKILL.md field manual and the field-tested skills roundup cover the ground.