If you've been shipping with Claude Code's SKILL.md format for any length of time, you've already noticed the distribution question forming. Where do skills live? Who decides which ones are good? What stops a single vendor from owning the format? Two models are emerging in parallel: open community catalogs that aggregate everything they can find, and walled marketplaces that vet, curate, and monetise.
This piece walks through both models honestly. Walled gardens have real strengths, and pretending otherwise gets you nowhere. But for a format as young as SKILL.md, where the conventions are still moving every month, the open catalog model has a structural advantage that's worth being explicit about. You'll come away knowing what each model is good for, how to spot the difference in practice, and why this particular catalog has made the choices it has.
Before arguing about which is better, it's worth pinning down what each one actually is, because the labels get used loosely.
A walled garden is a distribution channel where a single operator controls four things: which skills are admitted, how they're presented, who can pay whom for what, and what the format itself is allowed to become inside the channel. The Apple App Store is the canonical example outside our world. Inside it, you can imagine any number of plausible Claude Code skill marketplaces: a vendor-run store where every SKILL.md is reviewed before publication, billing is integrated, ranking is editorial, and the format may include vendor-specific frontmatter keys.
An open catalog is the opposite shape. The format is documented openly, anyone can publish a SKILL.md anywhere they like, and the catalog operator does discovery rather than admission. Skills are mined from the wider ecosystem — GitHub, GitLab, gists, package registries, sometimes direct submissions — and then sorted, ranked, classified and presented. Authors aren't asked permission. Authors aren't charged. The catalog itself can be forked or competed with at any time because the underlying data (where the skills live, what their SKILL.md files say) is public.
Both models can be high quality. Both can be low quality. The defining feature isn't quality, it's who holds the levers. In the walled garden, the operator holds them. In the open catalog, the levers either don't exist (admission) or are transparent and auditable (ranking).
A useful test: if the operator went out of business tomorrow, what happens to the skills? In a walled garden, the skills tend to be locked to that operator's distribution and metadata. In an open catalog, the skills were always living on their original homes; the catalog was a layer of organisation on top of them. The catalog can vanish and the ecosystem keeps moving.
One nuance worth getting right: a closed marketplace isn't automatically evil and an open catalog isn't automatically virtuous. There are openly-published, well-meaning catalogs that are operationally walled gardens because their admission process is opaque, and there are formally curated marketplaces that operate transparently enough to behave like open infrastructure. The shape that matters is who decides, on what criteria, with what recourse.
It's easy to be glib about closed marketplaces, especially if you've been burned by one. The honest case for them is actually strong, and any open-catalog operator who refuses to engage with it isn't worth taking seriously.
Quality vetting at scale. When every SKILL.md is read by a human before publication, the median quality of what a user sees is dramatically higher. That's not nothing. A new Claude Code user landing on a curated marketplace and installing the top-ranked skill in a category gets a working, well-documented, scope-disciplined skill on the first try. The same user landing on an open catalog has to learn to read the signals — which is fine for the seasoned reader, but a real cost for the newcomer.
A workable monetisation path for authors. An individual SKILL.md author has almost no way to get paid today. The walled garden offers a clean answer: list the skill, set a price, the platform handles billing and discovery, the author gets the net. This is the same value proposition the App Store made to indie iOS developers in 2008, and it's why so many of them happily traded thirty percent for distribution. For an author whose skill is genuinely worth charging for, the walled garden may be the only practical path to revenue.
Predictability and brand cohesion. A curated marketplace can guarantee that every skill follows a house style, ships with a video walkthrough, has been tested against the current Claude Code release, and won't pull in surprise dependencies. For an enterprise buyer making a procurement decision, that predictability is worth real money. "All the skills in this marketplace meet this bar" is a defensible statement in a way that "all the skills in this open catalog have been admitted via these rules" is not, even when the rules are public.
Concentrated trust signals. Brand cohesion compounds. A curated marketplace's reputation becomes a single signal that a user can rely on. They don't have to evaluate each skill from scratch; they trust the marketplace, and the marketplace stakes its reputation on each listing. This is the same machinery that makes hotel chains and supermarket private labels work. It's a real, useful service.
If your use case is procurement-driven, monetisation-sensitive, or aimed at a non-technical buyer, a walled garden is probably the right shape of distribution. Don't let anyone shame you out of that on principle. The principle is fit-for-purpose.
The flip side of the same argument: open catalogs win on a different axis, and the axis they win on matters more right now than it will in three years.
Discovery surface. An open catalog can mine the entire ecosystem and present it. A walled garden can only present what its operator has been pitched. The catalog you're reading this on currently has somewhere north of seventy thousand SKILL.md files, the overwhelming majority of which their authors never explicitly submitted anywhere. They just published their SKILL.md on GitHub, and the mining surface found them. A closed marketplace at the same age would likely have a few hundred listings, hand-shepherded.
Third-party innovation. When the admission gate is open, weird and interesting skills get published. Half of them are unusable; a small fraction are genuinely novel in ways no curator would have green-lit. The history of npm, PyPI, Chrome extensions, and Stack Overflow answers is the same story: the long tail of independent contributors produces the breakout patterns. Curators tend to fund what they already understand.
Faster iteration on the format itself. SKILL.md is six months old. The conventions are still moving. Anti-trigger sections, pricing tables in the body, sub-agent invocation patterns, allowed-tools frontmatter — none of these were obvious a quarter ago. An open catalog sees a new convention emerge in a few skills, ranks them well if their content quality holds up, and lets the convention spread. A walled garden has to convene a review committee, update its house style, and re-vet every skill against the new bar. The lag is real and it compounds.
No platform tax. Authors keep what they earn elsewhere, because the catalog doesn't intermediate billing. This is a constraint as much as a feature — it means the catalog can't fund itself off authors — but for an author who's giving a skill away anyway, it's the cleanest possible deal.
The downsides are real. Median quality is lower than a curated channel. Gaming risk is constant: anyone can publish, so someone will try to manipulate ranking. There's no clean monetisation path for authors who'd like one. Newcomers face a higher cognitive load.
The right framing is that open catalogs trade median quality for ceiling. You see more bad skills and more genuinely surprising ones. If you have the literacy to read the signals, the trade is heavily in your favour. If you don't, the curated channel does more of the reading for you.
Here's the structural argument for why open wins right now, even if you accept everything good about walled gardens.
Distribution channels can be sorted by what they optimise for. Walled gardens optimise for steady-state: a mature format with stable conventions, a known audience, predictable revenue, and slow-changing best practices. The App Store works because iOS as a developer platform changes slowly enough that a review process can keep up. Hardware specs are stable. Frameworks are versioned. The walled garden's review-and-vet machinery makes sense because the thing being reviewed isn't shifting underneath it.
Open catalogs optimise for explore: an immature format where conventions are still being discovered, where the audience is still being assembled, where the right shape of a good thing isn't fully known yet. The catalog's job in explore mode is to surface variation and let the ecosystem decide. You can't review a thing into existence; you have to see what people actually build and then figure out which patterns held up.
SKILL.md as a format is, today, deeply in explore mode. Examples: a year ago nobody had figured out that anti-trigger sections ("don't use this skill when...") materially improve agent behaviour. The skills that introduced this pattern weren't shipped by a curator — they were independent authors trying things, and the pattern propagated because other authors saw it working in the wild. Similarly with pricing/quota tables for API-driven skills, four-tier API funnel patterns, frontmatter conventions like allowed-tools and user-invokable. None of these would have been a curator's first guess. They emerged from the explore surface.
This is the structural reason open wins for SKILL.md right now. A walled garden in explore mode is shipping the wrong product: it's curating against a moving target, and by the time the curation catches up, the convention has moved again. An open catalog in explore mode is shipping exactly the right product: it surfaces every variant, lets the ones with content quality win on ranking, and watches the conventions converge in public.
In two or three years, when SKILL.md has stable conventions and a mature corpus of best practice, the steady-state arguments will get stronger. Walled gardens will be appropriate for many use cases — particularly enterprise procurement and paid-author monetisation. But right now, in mid-2026, the explore mode wins by structural argument, not by ideology.
If a walled garden launched today claiming to know the right conventions for SKILL.md, the polite reading is that they're calling their shot early. The less polite reading is that they're freezing the format prematurely to lock in their own house style.
The second structural argument is more boring and more important. Closed marketplaces depend on the platform owner's continued goodwill. Open catalogs don't.
Consider what "the platform owner" actually controls in a walled garden. They control who gets admitted (so they can de-list an author at any time, for any stated reason or none). They control how revenue flows (so they can change the cut at will, with the only recourse being to leave). They control discovery (so they can de-rank a skill they don't like). They control the format extensions inside their channel (so they can deprecate a key, mandate a new one, force a rewrite). They control the operating company itself (so the whole thing can be acquired, pivoted, or shut down).
Every one of these is a normal commercial reality. None of them is uniquely sinister. But each one is a single point of failure for the author, the reader, and the format.
An open catalog has none of these levers, because the catalog operator never had them in the first place. The skills live on their original homes (GitHub, GitLab, gists, personal sites). The format is documented in a public spec. The mining and ranking logic, if the catalog is honest about it, is published. If the catalog operator goes out of business, the underlying ecosystem is unchanged. Another catalog can spin up tomorrow on the same data.
This isn't a hypothetical. Look at any developer-format history of the last twenty years. Stack Overflow's role in the JavaScript ecosystem could be replaced if it had to be. CPAN's role for Perl was replaced by competing repositories without anyone losing their modules. Conversely, when a closed channel disappears — Yahoo Pipes, Parse, GitHub Atom, countless app stores — everything inside it disappears with it, and the recovery cost falls on the authors.
For a format you're betting your tooling on, this matters. If you're committing a year of work to writing high-quality SKILL.md files, you want to know that the distribution layer can fail without taking your work with it. Open catalogs give you that guarantee almost by definition. Walled gardens cannot, no matter how good their current intentions are.
The point isn't that any specific walled garden will fail or behave badly. The point is that you don't have to bet on either question. With an open catalog, your downside is bounded by the catalog being less useful. With a walled garden, your downside is bounded only by the platform owner's circumstances, which you don't control and can't predict.
The third structural argument is about the format itself. Closed marketplaces, over time, tend to extract value by extending the format in proprietary ways. This is the standard playbook and it's worth being explicit about how it works for SKILL.md specifically.
The vanilla SKILL.md format is open. A YAML frontmatter block with a small documented set of keys, a markdown body with conventional sections, a file living at ~/.claude/skills/<slug>/SKILL.md. Anyone can write one. Anyone can install one. There's no vendor required between the author and the reader.
A walled garden has a structural incentive to add extensions. Suppose the marketplace introduces a frontmatter key like marketplace-listing-id that handles billing and licensing inside the channel. Useful, on its face. Then a key like marketplace-required-runtime: 2.x that locks the skill to the marketplace's hosted Claude variant. Then a body convention where the install instructions reference the marketplace CLI rather than vanilla ~/.claude/skills/. Each step is individually defensible. The cumulative effect is that a SKILL.md written for the walled garden no longer works outside it without rewriting.
This isn't a prediction; it's the empirical pattern across every developer-format history of the last two decades. Browser-specific HTML, MSSQL-specific SQL, AWS-specific Terraform, Slack-specific Markdown variants. The flavour of Markdown alone exists in dozens of incompatible dialects because every major platform extended it. The vendor isn't usually being malicious; they're solving real problems for their channel. But the result is the same: skills written inside the channel can't easily leave it.
An open catalog has the opposite incentive. The catalog wants every SKILL.md to be portable, because the catalog's value is the breadth of the corpus. If skills start fragmenting into vendor-specific dialects, the catalog gets worse. The open catalog's interest aligns with the spec's interest. That's a rare and valuable alignment.
For an author, the practical answer is: write to the vanilla spec, publish your SKILL.md to a public home, let it be mined by whichever catalogs find it useful. If a walled garden wants to also list it, you can opt in — but the canonical version lives at the spec, not at any one channel. This costs you nothing and preserves every option you have.
For a reader, the practical answer is: when evaluating a skill, check whether its install instructions are vanilla (~/.claude/skills/<slug>/SKILL.md, plus any obvious dependencies) or whether they reference vendor-specific tooling. Vanilla is portable. Vendor-specific is the early warning signal of lock-in, even if the skill itself is currently free.
It's worth running the scenario out. What does the SKILL.md ecosystem look like in late 2027 under each model?
Walled-garden dominance. A small number of curated marketplaces hold the bulk of the discovery surface. Each has a house style, a review process, and an integrated billing path. Top authors are professionalised — there are people whose primary income comes from publishing SKILL.md files in one or two marketplaces. The median skill a user installs is high quality. The format has bifurcated: a public vanilla SKILL.md spec, and several marketplace-specific dialects with proprietary frontmatter and tooling. Authors who target a marketplace and authors who target the open spec are increasingly different populations. Innovation has slowed because new patterns have to clear a curator's bar, but the patterns that do clear it are well-engineered. Enterprise adoption is strong because procurement has a defensible target to point at. The risk surface is concentrated: if any single marketplace falters, a large fraction of the ecosystem's authors and consumers feel it.
Open-catalog dominance. A handful of major open catalogs aggregate the wider ecosystem, with overlapping but non-identical corpora. The format remains coherent because nobody has the lock-in incentive to fragment it. Authors publish wherever they like — personal repos, gists, awesome-lists, dev blogs — and rely on the catalogs to find them. The median skill is mediocre, but the ceiling is higher and the variation is richer. Tooling for evaluating skills (quality scores, install-retention signals, editorial picks) becomes a service layer on top of the open data, and competing tools can be tried side-by-side. Monetisation for individual authors is harder; some build services or consulting around their skills, others give them away as a portfolio. Innovation is fast because the explore surface is wide. Enterprise adoption is mixed: some firms build internal curation on top of the open catalog data, others wait for the open ecosystem to mature.
Neither scenario is dystopian. Both have real problems. The walled-garden scenario centralises quality and risk in the same hands. The open-catalog scenario distributes both, which is structurally healthier but harder to navigate as a new reader.
The honest forecast is that both will exist in parallel — the same way npm and the App Store both exist for different parts of the developer stack — and the interesting question is how much overlap there is in the middle. Authors will publish to both. Readers will pick based on use case. The format itself will be saved or broken depending on how seriously the major operators take portability.
The position this catalog has taken, deliberately, is that the open model serves the format better right now and that being honest about the trade-offs is more useful than partisan advocacy either way.
There's room for both models. The interesting question isn't which one is morally right, it's which one fits your situation as an author, a reader, or an operator.
If you're an author whose skill is genuinely worth charging for, who needs distribution infrastructure you can't build yourself, and who values predictable monetisation over format portability — a walled garden is probably your right move. You'll trade a cut for distribution, you'll write to the marketplace's house style, and you'll get a real audience faster than you would publishing into the open. Take the trade with your eyes open: keep a portable copy of your SKILL.md somewhere public, so that if the marketplace changes terms you have the option to leave. Don't let the marketplace's proprietary extensions creep into your canonical version.
If you're an author whose work is meant to be widely available, who already gives it away, who values surface area over revenue, and who wants to participate in the format's evolution — publish to the vanilla spec on a public home and let the open catalogs find you. You'll get reach faster and you'll keep all your options open.
If you're a reader with a clear, narrow use case and limited time to evaluate skills — a well-run walled garden saves you reading time. The curator has done some of the evaluation work; you trade variety for predictability. If you're a reader with broader interests, technical literacy, and the patience to read signals — the open catalog gives you a much wider field and the tools to sort it.
If you're an operator thinking about which model to run — the structural arguments above are real. Walled gardens are easier to monetise but harder to scale through a format's explore phase. Open catalogs are harder to monetise but cheaper to keep aligned with a fast-moving spec.
This catalog has chosen the open model intentionally, and the reasons are exactly the ones laid out in this piece. The format is too young to freeze. The discovery surface is too valuable to gate. The lock-in risk is too easy to slip into when revenue depends on authors. The platform risk is too one-sided for a closed channel to be a safe long-term home for the format. None of these arguments will hold forever; in three years they may not hold at all. But right now, in mid-2026, the case for open is structural and honest, and the case for walled is real but better suited to a different phase of the format's life.
If you're choosing between the two as a reader, the practical advice is to use both. Try a curated marketplace for the high-trust use case where you need to ship and don't want to evaluate. Use an open catalog for everything else, particularly when you're learning the format or hunting for patterns you haven't seen yet. The honest position isn't anti-walled-garden; it's pro-portability, pro-transparency, and pro-fit-for-purpose.
Found a bug or want a topic covered? Email [email protected] or open an issue via GitHub.
SKILL.md files, not affiliated with, endorsed by, or sponsored by Anthropic.