We just shipped 296 hand-authored Claude Code skill landing pages, with a guided-discovery surface and a public catalog dataset. Along the way we made about thirty real product decisions, half a dozen integrity catches, and learned some things that don't show up in the strategic-report PDFs about AI marketplaces. This is the honest version — what worked, what didn't survive contact with reality, and what we'd warn anyone building in this space to watch for.
The catalog is at skilllocker.ai/skills, the guided discovery is at skilllocker.ai/find, and the open dataset is at skilllocker.ai/skills.json. The lessons below are why it's shaped the way it is.
1. Version numbers must be earned — every single one
Early on we'd built a catalog where every skill displayed v1.3.0 and "3 iterations." Then we noticed something uncomfortable: a handful of skills only had two version files on disk (v1.0.0 and v1.3.0). The intermediate iterations existed conceptually but hadn't been preserved as separate artifacts. The version number was telling a story that the file system couldn't back up.
The right move wasn't to inflate the file count or quietly downgrade the displayed version. It was to be explicit about what v1.3.0 actually means — the semver minor version equals the iteration count, derived from the source file. If a skill says v1.3.0 and there are only two version files, the claim is "we did three iterations of conceptual change" — and we have to make sure that's true, not just convenient.
The discipline that follows: never mark a skill v1.x+1 unless a real Karpathy-loop iteration has happened. No cosmetic version bumps to make the catalog look maintained. The temptation is real and the right answer is never the convenient one.
2. There's a hard line between a "skill" and a "prompt"
This is the single biggest source of confusion in the Claude marketplace landscape. KissMySkills, the largest paid catalog right now, sells prompts at $12 and SKILL.md files at $39 — and the buyer often can't tell which they're getting. Aggregators bundle both under "skills." The free awesome-lists conflate them.
The actual line: a skill encodes judgment with a trigger description that lets Claude know when to apply it, automatically, without being asked. A prompt is a one-off instruction you paste in. A skill compounds across sessions; a prompt resets each time.
The buyer-facing implication: be explicit about what's in the box. Our pricing page says "Real SKILL.md files, not prompts in a folder" and the per-skill pages render the actual SKILL.md content (gated behind purchase) in their canonical format. Anyone copying the file into ~/.claude/skills/ gets a real skill that registers and triggers. Anyone copy-pasting a marketing "prompt" doesn't.
If you're shopping in this space: ask to see the SKILL.md format before paying. The frontmatter has name and description fields with trigger language. If those are absent, what you're buying is a paragraph someone calls a skill.
3. The Karpathy loop delivers more than three iterations of polish
The loop process — generate, test against held-out adversarial cases, identify the failure mode, patch, re-verify — does something subtler than "make the skill better." It surfaces the assumption you didn't know you were making.
Two concrete examples from the catalog:
-
The Anti-Slop Checker kept over-scoring ESL writing (non-native English speakers) as "AI-generated" because their patterns superficially resemble AI patterns: simple sentence structure, limited vocabulary range, occasional awkwardness. The held-out tests revealed the bias. The fix wasn't a tweak — it required adding context-calibration logic that the original framework didn't have. Without the loop, that bias would have shipped at scale.
-
The 3-Minute Email would default to "polite + thorough" framing for every email type, which is fine for some asks but disastrous for difficult conversations. Held-out cases (apologies, hard pushbacks, missed deadlines) revealed that the skill needed framework-matching logic before drafting — different email types use different structures, full stop.
The loop's real value isn't 30% better output. It's catching the failure modes you can't predict because they live in the assumptions you didn't know you were making.
4. Auto-extraction needs depersonalization
A practical lesson from the build: when you have ~300 SKILL.md files written in second-person voice ("you take vague prompts and rewrite them"), generating catalog metadata from them produces unreadable copy. "You take vague prompts and rewrite them" isn't a product description — it's an instruction.
We wrote a Python depersonalisation pass that handles the common patterns: "You verb..." → "Verbs..." (sentence start, third person), "You don't just X — you Y" → "Doesn't just X — it Y", "Your job/method/role is to X" → "X", mid-sentence "you" → "it", hyphenated verbs ("reverse-engineer" → "reverse-engineers"), and a guard so compound nouns like "thank-you" don't get mangled.
This isn't an interesting technical insight on its own. The lesson is: the format your skills are written in (instructions to the model) is not the format they should be sold in (descriptions to the buyer). If you're going to publish a catalog at scale, build the transformation pipeline. Otherwise you're either editing 300 descriptions by hand, or shipping copy that reads like onboarding instructions.
The depersonalisation pass is internal tooling rather than a separate shippable thing, but the concept is portable: any catalog generated from instructional source material needs voice transformation before it ships. If you're doing this at scale, build the pipeline.
5. Discovery is the universal unmet need — not catalog size
This was the single loudest signal from the competitive landscape research we did before building the discovery surface.
The numbers tell a clear story. Free aggregators index 1.4 million Claude skills indexed across the major sources. Anthropic ships 17 official examples. Corey Haines' MIT-licensed marketing skill pack is 46 files and gets cited everywhere. The most-installed Claude skill on claudemarketplaces.com is called find-skills — 1.5 million installs. Read that again. The single most-installed skill in the ecosystem is a skill whose job is to help people find other skills.
The lesson: in a market saturated with options, the product feature that matters most is the path through them. Catalog size is a signal of supply; navigation is a signal of design. We built /find as a three-step guided picker (role → task → 3 recommended skills) precisely because the 296-skill catalog without navigation is just 296 noise.
If you're shipping a catalog product — anything from skills to templates to tools — the navigation layer is more leveraged than the next 100 items. Spend time there.
6. Doubling bundle contents would break the price ladder
A counterintuitive one. The instinct says "more skills per bundle = more perceived value." The math says otherwise.
Current pricing has a clear $/skill ladder: single pillars at $1.81-$3.50 per skill, persona bundles at $1.77-$2.43 per skill, Full Stack ($299) at $1.01 per skill. The Full Stack tier is cheapest-per-skill by design — it's the volume play that rewards the buyer who wants everything.
Double the bundle skill count without raising the bundle price and the bundle $/skill drops to ~$0.88-$1.21 — below Full Stack. Now the volume tier has no value advantage. The upsell ladder collapses; the Full Stack price has to rise to restore the gap; bundle differentiation weakens because multiple bundles now share more pillars.
The lesson generalises: when expanding bundles, preserve the per-unit-cost gradient between tiers. If you add to bundles, raise the bundle price proportionally, or compensate by adding more to the top tier too. Otherwise you've devalued the up-sell by accident.
7. The free baseline shapes everything
This is the most uncomfortable lesson and probably the most important.
The competitive landscape for Claude skills is: free everywhere. Anthropic ships skills for free. Corey Haines' 46 marketing skills are MIT-licensed. Every awesome-list, every aggregator, every community GitHub repo serves the skills free. The dominant baseline is $0.
Selling skills in this market doesn't mean charging for what's free elsewhere. It means charging for what the free options can't deliver: curation, integrity discipline, hand-authored consistency, connected workflows across pillars, version management, lifetime updates. The product isn't "skills." The product is "the catalog you trust enough to stop searching."
Three implications that fall out of this:
-
Categories with strong free coverage are not where the moat is. Marketing skills are heavily covered by Corey Haines free; competing on volume there is a losing game. Our defensible pillars are the ones with weak external coverage: Course Creation, Health & Wellness, Fundraising, Voice-First, Agency Operations, Community Building, AI Memory, Learning. Lead with those.
-
The "what makes this worth paying for" claim must be specific and verifiable. "Higher quality" doesn't survive scrutiny. "Hand-authored by one human, looped through three adversarial-test cycles, version-tracked, with lifetime updates included" can be checked against the catalog. Either we're doing it or we're not.
-
Free assets are part of the offering, not a sales funnel concession. We ship five free skills (anti-slop-checker, three-minute-email, sixty-second-bio, content-idea-machine, meeting-debrief) and a free open-source plugin (context-window-awareness) because if our paid stuff is genuinely better, the free stuff proves the bar. If the free stuff isn't good, the paid stuff probably isn't either.
What the open dataset is for
The catalog metadata is now available as a free JSON dump at skilllocker.ai/skills.json. All skill names, descriptions, pillar mappings, bundle compositions, and prices — under a "public catalog metadata, free to read and build on" license. The actual SKILL.md content (the install file) stays gated behind purchase; the catalog itself is open.
If you're building a Claude tool directory, a comparison site, a research project on AI skill ecosystems, or you just want to see what 296 hand-authored skills look like organized — fetch it. Use it. Build on it. The only ask is that the source link goes back to skilllocker.ai if you publish anything derived from it.
What we still don't know
Honest list, for next time:
- Which skills actually get used. Without telemetry on the install side (deliberately so — we don't track what people do with skills after they download them), we have to infer from blog post traffic and search demand. The May 2026 GSC checkpoint is the first real data we'll have.
- Whether the pricing is right. Pre-traffic, every pricing experiment is theatre. We're holding prices steady until traffic lands and we can measure actual conversion.
- Whether the bundle taxonomy maps to how buyers actually self-identify. Six persona bundles is our guess. Real signal would tell us if "Coach & Consultant" should be two bundles, or if "Solopreneur" overlaps so much with "Freelancer" that they should merge.
Three months from now we'll know more about all of these. We'll publish what the data says, including where this retrospective got things wrong.
Until then: the catalog is at /skills, the discovery surface is at /find, the dataset is at /skills.json, and the free plugin is at github.com/miniminer-droid/context-window-awareness.
Try them on real work. That's the only test that matters.