The State of Claude Code Skills 2026

An even-handed field report on the agent-skills ecosystem — what exists, what it costs, what's good, and what's hype. I run Skill Locker, a paid catalogue of Claude Code skills, so I have a stake in these conclusions; I've labelled our own data as ours throughout and tried to make this useful even if you never buy anything from me.

TL;DR

• The category went from zero to mainstream in about 18 months. Anthropic introduced the SKILL.md format in October 2025, opened it as a cross-agent standard in December 2025, and reportedly launched a public Skills Marketplace (~600 skills at launch) by May 2026 (reported by Anthropic; raw counts may include overlaps). For context, npm took roughly a decade to reach 350k packages; agent-skill registries claim five- and six-figure counts within months — though those raw counts are noisy (see below).
• Nobody actually knows how big the ecosystem is. Published catalog sizes range from ~600 (Anthropic's marketplace) to ~7,000 (SkillHub) to ~30,000+ (SkillsMP) to a community registry that describes itself as backing 70,000+ SKILL.md files — but these overlap heavily, mix skills with agents/prompts/MCP servers, and use wildly different quality bars. There is no reliable ecosystem-wide count, revenue figure, or market-share data (no authoritative source exists).
• The demand context is strong; the paid-skill demand is unproven. Claude Code's weekly active users roughly doubled in early 2026 and reportedly surpassed ~2M by May 2026, on an annualized revenue run-rate around $2.5B by February 2026 (figures circulated by Anthropic and secondary trackers, not independently audited). But surveyed willingness to pay for generative-AI tools is modest (~37% in one 2025 survey, with creative tools a notable exception at ~45%), and most skills in circulation are free and open-source. Whether people pay for skills specifically is still open.
• The free tier dominates discovery; paid marketplaces are small and fragmented. Open-source awesome-* lists and free registries hold the vast majority of skills. Paid players (Agensi at typically $5–15/skill, ClaudSkills at $9/mo or $149 lifetime, KissMySkills at an inferred ~$39/skill) are niche by comparison.
• Most "skills" are prompts in a folder. The gap between a real skill (scoped trigger, progressive disclosure, evals, negative test cases) and a glorified prompt is enormous — and most headline counts are padded with the latter.
• 2026 looks like a consolidation year, not a land-grab. Gartner positions agentic AI past the early-demo phase: the frontier has shifted from "can it do this?" to "does it do this reliably and safely?" Consolidation toward a few trusted hubs is plausible, with value moving toward curation, verticalization, and trust signals rather than raw catalog size.

1. What "skills" are, and why 2026 is the inflection

A skill is a folder containing a SKILL.md file: YAML frontmatter (a name and a description) plus markdown instructions, optionally with reference files and scripts. The agent reads the description to decide whether the skill is relevant to the task, then loads the full instructions only when needed — a design Anthropic calls progressive disclosure (Anthropic docs). Only name and description are strictly required, which is exactly why the format spread so fast and why quality varies so much.

Three things converged to make 2026 the inflection point:

1. A standard. Anthropic shipped SKILL.md (Oct 2025) and opened it as a cross-agent standard (Dec 2025). The format is now supported by all major AI coding agents, including Claude Code, OpenAI Codex, Cursor, Gemini CLI, and GitHub Copilot (agensi.io). Portability, not lock-in, is the strategy — the same playbook Anthropic ran with MCP.
2. A distribution channel. Anthropic's in-product plugin directory catalogs 55+ curated plugins (code.claude.com), with community marketplaces adding more; separately, the public Skills Marketplace launched ~May 1, 2026 with roughly 600 skills, one-command install, and a 15% cut on paid skills (siliconangle, venturebeat). Centralized discovery, version pinning, and permission controls followed. Anthropic also launched an enterprise-facing Claude Marketplace (March 6, 2026) with six vetted partners — GitLab, Harvey, Lovable, Replit, Rogo, Snowflake — and notably takes no commission on those enterprise purchases (siliconangle, venturebeat).
3. A non-developer audience. Anthropic's "Cowork" interface (early Jan 2026) explicitly targets non-programmers, and people are using Claude Code for SQL, landing pages, content research, and ops without writing code (xda-developers). This broadens who a skill is for — and reshapes which categories matter.

2. The landscape: who's who

Four distinct layers, often conflated in headline counts:

Official (Anthropic). The plugin directory (55+ curated) and the public Skills Marketplace (~600 at launch, mostly free, 15% cut on paid) emphasize free distribution, curation, and commercial-partner participation (the Agent Skills directory partners include Atlassian, Canva, Cloudflare, Figma, Notion, Ramp, and Sentry). Plus the enterprise Claude Marketplace for partners. Anthropic is competing on trust and integration, not catalog size.

Open-source awesome-* lists and registries (the bulk). This is where most skills live, and it's free:

• VoltAgent/awesome-agent-skills: 1,000+ cross-agent skills.
• alirezarezvani/claude-skills: 337 skills; Orchestra-Research/AI-Research-SKILLs: 87 skills across 22 categories.
• ClaudeSkills.info: 658+ browseable skills. The same project describes itself as the community-curated registry of 70,000+ SKILL.md files — that larger figure is the registry's own self-description (a scrape-scale claim), not independently verified, and should be treated as such.

Aggregators (raw scale, low curation). SkillHub (~7,000 skills, positioned as curated) and SkillsMP (~30,000 entries — but mixing skills, agents, prompts, and MCP servers; estimate, per each platform's own landing pages as of mid-2026). Big numbers, inconsistent quality, and counts that shift over time.

Paid marketplaces (small, fragmented). Agensi (curated, 80/20 creator split, typically $5–15/skill with specialized skills reaching $25, Stripe Connect payouts — verified against agensi.io's public site as of June 2026); ClaudSkills ($9/mo or $149 lifetime, with a quality score and one-click install — verified against ClaudSkills' public pricing); KissMySkills (positioned as premium; ~$39/skill is inferred from competitor analysis — pricing not publicly disclosed). By Q2 2026 there were ~8 agent-skill marketplaces, up from essentially zero in 2024 (agensi.io).

A useful sense of scale: the adjacent MCP ecosystem — same vendor, same era — had roughly 17,468 indexed servers in a Q1 2026 census, and MCP SDK downloads reportedly grew from ~2M/month (Nov 2024) to ~97M/month (Mar 2026) as OpenAI, Microsoft, and AWS adopted it (per a single analysis at digitalapplied.com; growth figures not independently audited). The most useful number from that same source: only ~12.9% of MCP servers score "high trust" (70+/100) under one Q1 2026 audit — there is no authoritative trust-score standard for MCPs, so read it as one analysis, not consensus. The lesson generalizes regardless: scale tends to outrun quality, and skills face the same risk curve — though MCP is not skills, and the parallel is a risk, not a proven equivalence.

3. Categories and gaps

The honest version: there is no authoritative taxonomy of skill categories or their relative sizes. What we can observe is where installs cluster and where there's visible thinness.

Where skills concentrate. Developer workflows dominate. A TechTimes write-up of Anthropic's framework reports nine workflow categories and ranks verification as the single highest-impact type (TechTimes, on Anthropic's framework — verify against the TechTimes article and Anthropic's docs). On Agensi, the top-installed skills are overwhelmingly engineering- and content-adjacent: code-reviewer (116 installs), git-commit-writer (65), readme-generator (49), pr-description-writer (36), env-doctor (30), changelog-generator (27), seo-optimizer (21), humanize-writing (16) (agensi.io; one platform's data, not the ecosystem's). Marketing is a fast-growing adjacent cluster — PPC automation, creative optimization, audience research, competitor analysis — and GEO (Generative Engine Optimization) features have become standard in SEO skills as of 2026 (get-ryze.ai, firecrawl.dev).

Where it's thin (one lens — our own catalog). To make the gap concrete I'll use the taxonomy from Skill Locker, the catalog I run (full disclosure in the methodology note). These are our catalog's numbers, not ecosystem figures — citable only as a description of one hand-authored catalog: 296 skills across 33 pillars. Plotting those pillars against where the open ecosystem clusters, the consistently underserved zones are the non-developer, knowledge-worker categories: Learning & Growth, Course Creation, Voice-First, AI Memory, Health & Wellness, Agency Operations, Fundraising, Community Building, and Personal Brand — roughly 8–10 of our 33 pillars, and a poor match for what the developer-centric awesome-* lists and install charts cover.

A caveat on that gap: the supply appears heavily developer-built based on visible install charts and public lists — but that may partly reflect where skills are visible and counted rather than the true distribution of what exists. With that hedge, the macro signal still points the same way: Anthropic is itself pushing past horizontal developer skills into verticals — financial-services agent templates for pitchbook building, KYC screening, and month-end close shipped May 2026 (anthropic.com). The gap, in one line: the visible supply skews toward developers; the demand is broadening to everyone else.

4. Pricing reality

Start with the uncomfortable fact: paid-skill demand is unproven. One 2025 survey put willingness to pay for any generative-AI tool at ~37% (Suzy, 2025; survey data, not freely verifiable) — with a notable exception for creative tools, where willingness rises to ~45% for music/image generation (Statista, 2025). There is no Claude-skills-specific willingness-to-pay study. The free open-source layer is the default, and it sets a hard reference price of zero.

What we do have is pricing behavior from the platforms and from comparable digital-asset markets:

Two reference points from adjacent markets are instructive (both from a single analysis at insightsetter.com; not independently verified):

• Prompt marketplaces (PromptBase) sit at roughly $1.99–$9.99 per item — the atomic, commodity floor.
• Gumroad digital products are reported to convert best in the $30–49 band (per insightsetter.com analysis; the ~28%-better-than-sub-$10 figure is one analysis, not consensus), with software averaging ~$39.95 and courses ~$95.74. This suggests packaged, multi-step assets may command meaningfully more than per-prompt pricing.

And a cautionary one: GPT Store revenue-share has been reported to average ~$0.03/conversation — implying ~33,000 quality conversations/month to clear $1,000, with anecdotal reports of most creators earning roughly $100–500/month (digitalapplied.com; anecdotal, not a measured statistic). Revenue-share-on-usage has been a weak monetization model; one-time/bundle purchases have produced higher lifetime value for comparable creators.

For full disclosure, our own pricing — again, one data point, the author's project — runs $29–49 per pillar, $99–149 for persona bundles, $299 for the full 296-skill stack. That sits deliberately in the Gumroad "packaged digital product" band rather than the per-prompt commodity band. Whether that band holds for skills specifically is exactly what the market hasn't proven yet. Anyone who tells you paid skills are a validated business is ahead of the data.

5. What makes a skill actually good

This is the part the headline counts obscure. A 30,000-entry catalog and a 30-skill catalog can contain roughly the same number of good skills. An opinionated, source-grounded quality bar:

1. Single responsibility. One skill does one thing well. Red flags: a description with "and" joining two distinct operations, or unrelated output fields. The responsibility can be substantial ("extract all action items from a meeting transcript") but must be singular (mindstudio.ai).
2. A description that encodes when to trigger, not just what it does. The description (≤1024 chars, no XML tags) is what the agent reads to decide relevance. "Process the meeting transcript" is too vague; "Extract action items from meeting transcripts and return structured JSON" triggers correctly (Anthropic docs).
3. Progressive disclosure. Don't dump catalogs/policies into the prompt — retrieve only what's needed at execution time. Keep SKILL.md under ~500 lines; push heavy reference material to a references/ subfolder (Anthropic docs; digitalapplied.com).
4. Gotchas over restatement. The highest-signal content is the non-obvious stuff that pulls the agent out of its defaults — edge cases, team rules, constraints. Restating default behavior is pure noise (Anthropic engineer, via Medium).
5. Evals before docs — including negative cases. Run the agent without the skill first, document the real failures, then write 3–5 representative test cases covering should-trigger, should-not-trigger, and edge cases. The negative cases are the most revealing — a skill that correctly says "not my job" is more valuable than one that forces itself everywhere (Anthropic docs; mindstudio.ai).
6. Iterate against a measured baseline. A binary-assertion eval loop — measure pass rate (often plateauing 85–95%), let the agent analyze failures and refine, re-test — yields a handful of meaningful improvements per cycle. Run each case multiple times to smooth non-determinism (zerocopy.blog).
7. Know what isn't a skill. "Always do X" standing rules belong in CLAUDE.md; 12-step procedures belong in a skill. Mixing them is a top anti-pattern, and unscoped skills surface every session, bloating context (CLAUDE.md reportedly starts getting ignored past ~200 lines) (digitalapplied.com; claudefa.st).

The dominant failure mode in the wild is skill sprawl: teams accumulate 40–50 skills in six months and actively use ~6, because there's no governance on scope, metadata, or lifecycle (digitalapplied.com). More skills is not better. Tested skills are better.

One data point on why testing matters commercially, not just technically: workers with AI skills are reported to earn wage premiums of up to ~56%, and analyses project net-positive job impact when organizations intentionally redesign work — 170M roles created vs. 92M displaced (World Economic Forum, Future of Jobs projections). The signal: the value is migrating to demonstrable capability, not raw access.

A note on my own claim, for honesty: our catalog versions every skill (v1.x, with changelogs) and runs a 3-round eval loop with held-out test sets. That's our methodology — not an industry standard. The full skill metadata is published openly so you can check the claim yourself rather than take it on faith (links in the methodology note, both confirmed live at the time of writing). Treat any vendor's quality claims — including mine — as something to verify against published evidence.

6. Predictions for the rest of 2026 (opinion — clearly labeled, not fact)

These are reasoned bets, not facts:

• Consolidation is plausible — but not inevitable. If 3–4 marketplaces emerge as trusted hubs, network effects and the cost of curation could thin the field from the current ~8. But horizontal aggregation (the 30k+ catalogs) could also persist on sheer convenience. I lean toward consolidation; I wouldn't bet the house on it.
• Trust scores become table stakes. MCP's reported ~12.9%-high-trust problem will likely recur for skills. Expect quality scores, verified-publisher badges, and eval transparency to become standard discovery filters — and a genuine differentiator.
• Verticalization may beat horizontal volume. Following Anthropic's finance templates, domain bundles (legal, HR, healthcare, agency ops) could outperform generic horizontal catalogs. Disclosure of obvious self-interest: this prediction aligns directly with Skill Locker's own taxonomy (Agency Operations, Personal Brand, Fundraising, etc.), where open-ecosystem supply is thin — so weight it as a forecast I have a stake in, not neutral analysis.
• "Routines" (scheduled/event-triggered skill executions) become a watched feature. Marketplace operators identify routines-plus-skills as the most bullish 2026 bet (firecrawl.dev); whether they become the most consequential feature is my speculation, not theirs.
• Free stays the default; paid survives in B2B and specialist niches. Given modest consumer WTP and a deep free layer, sustainable paid demand likely concentrates in enterprise licensing and specialist verticals, not consumer-direct.
• The honest unknown: I don't expect a credible ecosystem-wide market-size figure to emerge this year. Overlapping catalogs and incompatible quality bars make a real census very hard. Distrust any precise number you see — including, eventually, this one.

About this report / methodology

This report synthesizes publicly cited sources on the Claude Code / agent-skills ecosystem as of mid-2026, with inline attributions. Ecosystem-wide figures are estimates or single-source claims unless context indicates otherwise, and several key numbers (total skill count, market size, market share) have no reliable public source — I've said so explicitly rather than inventing precision. Catalog sizes from competing platforms are reported as those platforms describe them and are not independently audited; counts overlap, mix content types, and change over time, so they are not directly comparable. Install counts (Agensi) reflect one platform's data, not the ecosystem.

Disclosure: I run Skill Locker (skilllocker.ai), a paid catalog of 296 hand-authored Claude Code skills across 33 pillars. Where I've used our catalog — the 33-pillar taxonomy, the underserved-category analysis, our pricing, our eval methodology — I've labeled it as our own data, not ecosystem fact. I have an obvious commercial interest in the conclusion that curation and testing beat raw volume; weigh that accordingly. In the spirit of being checkable rather than asking for trust, the full metadata for all our skills is published free for research, tooling, and comparison at skilllocker.ai/skills.json (publicly accessible — confirmed live at publication), with five skills fully open-source under MIT at github.com/miniminer-droid/skill-locker-free (also confirmed live). Use them to fact-check this report, or to build something better.

If you remember one thing: the skill ecosystem's likely binding constraint in 2026 isn't supply — it's trust. Plenty of operators are betting on breadth instead, and they might be right. But my money is on whoever makes quality legible.