What changed in the catalog

The 42 agent-vs-agent comparisons derived their dimension table, overlap counts, verdict and FAQ inside the page component, so a markdown twin would have been a second derivation of the same numbers. The derivation moved into a data module instead — one source, read by both — and the twins followed. Every page type this site renders now publishes a machine-readable form.

• New surfaceAgent comparisons derive once and publish a twin

  buildAgentComparison() now computes the dimensions, the verified counts and their overlap, the verdict and the FAQ, and both the page and the twin read what it returns. Agreement is structural rather than something a future edit has to remember. The refactor was held to changing nothing a reader sees: all 42 pages came out of the build with identical visible text and identical JSON-LD, compared against the previous build rather than assumed. Sitemap 952 to 994.

  /compare/claude-code-vs-cursor →

Comparisons, collections and the category x agent cuts were the last page types available only as HTML. 43 more twins, which completes the set for everything this site renders from data.

• New surfaceMarkdown twins for head-to-heads, collections and category x agent cuts

  A comparison twin leads with the verdict and carries the full dimension table; a collection twin carries the ranked picks; a cut twin carries both halves — the Skills verified on that agent AND the Skills in the category that are not. That second half is deliberate. Saying which Skills have not been tested on an agent is the claim competitors never make, and a twin that dropped it would publish only the flattering half of the page. The one page type still without a twin is the agent-vs-agent comparison, and for a reason worth stating: its dimension table is derived inside the page component rather than from a data file, so a twin would be a second derivation of the same facts. Two derivations drift, and a drifted twin is worse than none — that needs an extraction first, not a quick export.

  /top-browser-automation-skills-for-cursor →

Skill, category and per-agent pages have published a markdown twin for a while. Guides did not — and the first measurement of which queries actually cite this site found both hits resolved to a guide, with the engine lifting the lead sentence close to verbatim. 28 guides now publish a twin.

• New surfaceMarkdown twins for the agent, explainer, job and launch guide clusters

  Every twin leads with the same citable first sentence the page does, then the comparison table as a real markdown table, the steps as an ordered list, the prose, the FAQ and the outbound links — all from the same data the page renders, so the two cannot drift. The older hand-written guides keep their prose in JSX and are deliberately left out: a twin there would mean a second copy of every paragraph, two copies of a sentence drift apart, and a drifted twin is worse than none. The <link rel="alternate"> appears only on the 28 pages that actually have one, so nothing advertises a twin that does not exist, and a request for a twin that was never built 404s rather than guessing.

  /guides/how-to-buy-a-domain-with-an-ai-agent →
• VerifiedFirst measurement of whether answer engines cite this site: 2 of 10

  The scoreboard metric that has read "unmeasured" since August. Ten "best agent skill for X" probes through one search backend — one engine of the three the target names, so a floor rather than the full picture. Cited on deploying a website and on document processing, and in both the engine quoted our prose close to verbatim, which is the whole bet behind leading every page with the answer. Of the eight misses, four are pages one to three days old and prove nothing yet; the other four are established queries where competitors rank with one dedicated page per narrow query. Worth recording that the domain query we lost returned the confident claim that Claude cannot buy a domain and you must finish at a registrar — the exact thing our guide corrects with two verified skills.

  /guides/best-skills-for-domains-and-dns →

The per-agent install tabs rendered only the active panel, so a Skill verified on five agents published one command in its HTML and hid the other four behind a click. The .md and .json twins carried the full set the whole time, which made the page itself the thinnest surface this site publishes.

• New surfaceEvery install panel is now in the HTML, inactive ones hidden rather than absent

  firebase-hosting is verified on Claude Code, Cursor, Codex, Gemini CLI and Antigravity, with a different command for each. Only the Claude Code one reached the rendered page; the rest existed solely after a click, so a crawler indexing the page — or anyone reading it without JavaScript — got one command in five. Rendering every panel and marking inactive ones `hidden` is the accessible tab pattern anyway, and it came with the ARIA wiring that was missing: tabs now carry ids and aria-controls, panels carry role and aria-labelledby, and the copy button names the agent it copies for. Verified across all 206 skill pages: every tab has a rendered panel, and no duplicate element ids.

  /skill/firebase-hosting →

The composite score rendered through a count-up animation that starts at 0. Server-rendered HTML is what a crawler reads, and crawlers execute no JavaScript — so GPTBot, ClaudeBot and PerplexityBot read "Score 0" out of all 206 skill pages, for the number this site is built around. Also adds the WordPress project’s own skill set, and guides for migrating a site and running WordPress with an agent.

• New surfaceThe score is now correct in the HTML, not just after hydration

  The badge used an animated counter initialised to 0 and eased up to the real value on mount. For a reader with JavaScript that looks like a nice touch. For the audience this site is written for it was a wrong number: AI crawlers execute no JavaScript, which is the reason the static export exists, so every crawl of every skill page saw a score of 0. The value was correct in the badge’s data attribute and in the JSON-LD the whole time, which is exactly why nothing caught it — the page was right in every machine-readable place except the one a crawler reads as text. The counter now initialises to the real value and animates only when the value changes, which on a static page means it never fires. That is the right trade: the animation was decoration, the score is the argument.

  /about/methodology →
• Addedwordpress-agent-skills — the WordPress project’s own 18-skill set

  GPL-2.0, 2,200 stars, first-party, and a real omission until now. It exists because agents reliably generate pre-Gutenberg patterns, skip block deprecations in a way that produces "Invalid block" errors on existing posts, and miss the security conventions around the settings API. Covers blocks, block themes, plugins, the REST, Interactivity and Abilities APIs, WP-CLI and multisite, performance profiling, PHPStan and Playground. The project discloses that the skills were generated from its own documentation with a model and then reviewed by contributors, and labels them v1 — a more honest provenance statement than most first-party sets make.

  /skill/wordpress-agent-skills →
• New surfaceTwo guides for jobs the site could not answer at all

  Migrating a live site to a new host, and building then running WordPress with an agent. Both were unanswerable here a week ago and are now backed by verified skills. The migration guide leads with the sequence rather than the tooling, because most bad migrations are not import failures — they are a DNS change made before anyone checked the copy. The WordPress guide is deliberately split down the middle: the WordPress project’s set teaches an agent to build, a host’s set keeps the running site patched, and it says plainly that the operating half is one publisher deep and only reaches sites hosted there.

  /guides/how-to-run-a-wordpress-site-with-an-ai-agent →

Hostinger ships two skill repositories, not one, and the second is the more useful of the pair: outcome-shaped skills for migrating a live site off another host, connecting a domain registered anywhere, and keeping WordPress patched. Cataloguing it also refuted something these guides had been asserting since 2026-09-17 — that no skill spans the seam between a domain at one registrar and a site at another host.

• Addedhostinger-api-mcp-skills — the jobs, where the other repo has the endpoints

  Seven skills on Hostinger’s MCP server, MIT, 156 stars and committed the day it was catalogued: deploy-to-hosting with auto-deploy from GitHub, migrate-to-hosting for moving a WordPress or PHP site off another host without downtime, connect-domain, maintain-wordpress, audit-hosting, troubleshoot-website and a headless build skill. It covers two jobs nothing else in the catalogue does at all — migrating an existing live site, and WordPress maintenance. Worth noting how carefully these are written: the migration skill keeps the old host serving until the copy is proven and moves DNS last, the audit skill is read-only end to end, and the WordPress one updates a site at a time and confirms each still loads. Install executed, not transcribed.

  /skill/hostinger-api-mcp-skills →
• Verified"No skill spans the registrar-to-host seam" was too strong, in four places

  The domains guides said that connecting a domain bought at one company to a site hosted at another is the step no skill covers, because no company holds credentials on both sides. The second half is true and the first half was not. Hostinger’s connect-domain accepts a domain registered anywhere: it verifies ownership with a TXT record, attaches the domain, installs the certificate, enables the HTTPS redirect, and carries existing MX, SPF and DKIM records through the change rather than overwriting them. Everything on the hosting side is automated. What remains is the single A or CNAME at the other registrar, which it names for you and cannot write. All four places now say that instead — the narrower claim is also the more useful one, because it tells a reader exactly how much is left to do.

  /guides/connect-a-custom-domain-to-an-agent-deployed-site →
• New surfaceHowTo structured data on the four how-to guides in the domains cluster

  Every deploy guide emits HowTo markup; the domains-and-hosting cluster shipped without it, including the four guides whose titles begin "How to". Those are the queries an answer engine resolves from steps. The shared guide layout now renders an ordered step list and emits HowTo from that same array, so the markup describes what is on the page rather than making a claim about it — the condition Google puts on this markup, and the reason to render rather than only declare. The steps are the commands verified against these skills, not a generic recipe. The two guides in the cluster that rank and compare rather than instruct get no steps and no markup.

  /guides/how-to-buy-a-domain-with-an-ai-agent →

The entry announcing three new skill-vs-skill comparisons linked one of them at the agent-comparison path. This site has two comparison surfaces that differ by one path segment, and that link used the wrong one — so it 404’d from the moment it was written. It is fixed, and the build now fails on any editorial link that resolves to no published page.

• New surfaceThe build now rejects an editorial link that goes nowhere

  Editorial data carries 91 hand-written hrefs across this changelog, the explainer guides and the domains cluster. They are plain strings, so a path that resolves to nothing compiles cleanly, ships, and 404s whoever clicks it — the link is valid TypeScript, valid HTML, and present in the rendered page, so nothing objects. A new check derives the set of paths the site actually publishes, from the same data the sitemap is built from, and fails the build on any editorial link outside it. Where a /compare/ path would resolve under /compare-skills/ it names that as the fix, because the two surfaces differing by one segment is exactly the mistake a human re-reading the diff does not catch. It also enforces the rule that editorial links use the public category and agent URLs rather than the internal paths those rewrite from.

  /compare-skills/chrome-devtools-vs-playwright-cli →
• New surfaceCategory x agent pages and collections were each reachable from one place

  An audit of internal links across all 388 pages found no orphans, but a long tail on a single inbound link. The category x agent cuts were linked only from their category page, despite being the pages that answer "which of these work on Cursor" — and a cut is literally "Top <category> Skills for <agent>", so the agent page is its other natural hub and now carries them. Collections were reachable only from the collections index, which also meant a reader on a skill page never learned it was an editorial pick; skill pages now list the collections they appear in, read back from the same membership computation rather than a second list that could disagree. Minimum inbound links: cuts 1 to 2, collections 1 to 7.

  /collections →

The hostinger-vs-zeabur comparison exists because both were added the same day and a reader picking between them had nothing to go on. Researching it caught four errors in the entries we had just shipped — a wrong skill count and a self-contradicting "the only platform that…" in three places — so this release is the page plus the corrections it forced.

• New surfacehostinger-agent-skills vs zeabur-agent-skills, across ten dimensions

  The obvious framing was wrong and got thrown out: these are not a registrar versus a container platform, because both register domains and both manage DNS. The axis that actually decides is what else has to happen in the same session, and whether the zone can survive a mistake. Hostinger is the only one of the pair with a validate-before-apply dry run, zone snapshots and a restore; Zeabur is the only one whose skills reliably fire, with 25 of its 31 loadable descriptions quoting the words a user would type against 0 of Hostinger’s 8. That count was measured over every description in both repos, not estimated.

  /compare-skills/hostinger-agent-skills-vs-zeabur-agent-skills →
• VerifiedZeabur ships 32 skill directories, not 24 — and one of them cannot load

  The count came from the repo README, which lists 23 and names zeabur-domain-registrant, a skill that does not exist as a directory; its commands live inside zeabur-domain-register. The real tree has 32 directories. Of those, 31 have a reachable SKILL.md: zeabur-template-deploy ships its file as skill.md in lowercase, so on a case-sensitive filesystem no loader sees it. Confirmed rather than inferred — asking the skills CLI to install it by name fails and the skill is absent from the list of available skills it prints back. Reported here because the same failure will hit anyone who follows that README.

  /skill/zeabur-agent-skills →
• VerifiedA "the only platform that…" claim that was false in three places

  Both new entries were written the same day, and three surfaces called Zeabur "the only catalogued platform that is both the host and the registrar". Hostinger is one too — which the very next sentence of the launch guide said out loud, describing it as "a registrar that also sells the hosting". A superlative and its own counterexample, two lines apart. All three are rewritten to say what is actually true: two platforms are both, and they approach it from opposite ends. Superlatives are now the first thing to re-check before a page ships.

  /guides/launch-a-website-with-ai-agent-skills →

Skill pages have shipped markdown and JSON twins since launch. The pages an answer engine actually lands on for "best skill for X" are the category and per-agent rankings, and those were HTML only — so a crawler had to parse a rendered document to reach a ranking the site already holds as data. All 36 of them now publish both. The JSON Schema every one of these documents declares was a 404; it is now served and every published document validates against it.

• New surface72 machine-readable twins: 29 category rankings and 7 per-agent rankings, each as .md and .json

  A category page is the site's answer to "what is the best skill for X", and until now that answer was only available wrapped in nav, cards and a footer. /top-skills-for-<slug>.md and .json now carry the same ranking with no chrome: the citable lead, every entry with its score and provenance, the install command for the top five, and the FAQ answers. The per-agent twins go one better and carry that agent's own install command per Skill rather than the catalogue default, which is the one thing a generic export cannot give an agent working on a particular surface. Both twins and the HTML read one source, so they cannot drift. Linked from each page with rel="alternate", listed in the sitemap at a lower priority than the HTML they mirror, and announced in llms.txt.

  /top-skills-for-web-hosting-deploys →
• New surfaceThe JSON Schema every catalog document advertised was a 404

  Every /skill/<slug>.json declared $schema: /schemas/skill.v1.json and /skills.json declared /schemas/catalog.v1.json. Neither resolved. That is worse than declaring nothing: a consumer that dereferences the URL to validate gets an error page, and a reader who clicks it learns the declaration was decorative. Both are now served, documenting the three published shapes — the whole catalog, a category ranking, a per-agent ranking — and the skill entry, including the fields whose meaning is load-bearing: what a provenance tier asserts, that compatibility lists verified agents only, and that verifiedAt is the date the install was actually run. Every document the site publishes was checked against them.

The site could deploy a site to six hosts and then had nothing to say about the two steps on either side: registering the name, and pointing DNS at the thing you just shipped. Two first-party skills that do both are added, and a six-page cluster covers the launch chain end to end. The de-listed domains category stays de-listed, and this entry says why.

• Addedhostinger-agent-skills — the registrar surface, with DNS you can roll back

  Hostinger's first-party MIT-licensed plugin, eight skills: domains (registration, availability, WHOIS profiles, nameservers, forwarding, transfer locks, privacy), dns (zone records with snapshots, validation and reset-to-defaults), hosting (sites, free subdomains, databases, WordPress, Node.js apps), horizons (their AI website builder), plus vps, billing, ecommerce and reach. The snapshot-and-restore path is the strongest single argument in this area and nothing else catalogued has it. Install was executed on 2026-09-17, not transcribed: npx skills add hostinger/hostinger-agent-skills --skill domains --yes resolves and writes a universal .agents/skills/ folder with a Claude Code symlink.

  /skill/hostinger-agent-skills →
• Addedzeabur-agent-skills — a registrar that is also a container platform, in one CLI

  Zeabur's first-party MIT plugin, 32 skill directories of which 31 load. The deploy half was already well covered by others; what is new here is that the same CLI registers domains outright across nine TLDs (.com .net .org .io .dev .app .co .me .xyz), holds the WHOIS registrant profile, handles renewal and ICANN verification, and writes DNS records across A, AAAA, CNAME, MX, TXT, SRV, CAA and NS. That makes it the shortest path from searching a name to a live site on it. The repo documents Codex alongside Claude Code, and both install paths were run on 2026-09-17.

  /skill/zeabur-agent-skills →
• New surfaceSix guides on domains, DNS and where a site actually lives

  A pillar on the four-step launch chain (build, deploy, register, point DNS), plus how to buy a domain with an agent, how to manage DNS with one, how to connect a custom domain to a site an agent deployed, the best skills for domains and DNS, and where to host a site an agent built. Each one names the skill that does the step and, more usefully, where the chain breaks: no skill spans the seam between a domain bought at one company and a site hosted at another, because no company holds credentials on both sides.

  /guides/launch-a-website-with-ai-agent-skills →
• VerifiedThe domains-and-DNS category stays de-listed, and the reason changed

  It was de-listed in 2026-07 because nothing cleared the bar. Two skills now do, which is a real change and is why the guides above exist — and it is still not a category. Two entries cannot be a ranking, and padding to ten would mean shipping exactly the entries the de-listing removed. A full search of SKILL.md files mentioning domains or DNS returns over a thousand hits that are availability checkers, WHOIS lookups, OSINT reconnaissance tooling, or DNS editors for a zone you already control. One well-regarded candidate routes every operation through its author's own checkout — its instructions read "use ~/Projects/manager as truth" and call scripts at paths that exist on one laptop — so it works for exactly one person and was rejected. The underlying constraint is structural: registering a domain needs a registrar API and a billing relationship, which a community skill cannot have. The category returns when ranking says something a list of two cannot.

  /guides/best-skills-for-domains-and-dns →

Six FAQ answers write file paths and Skill names in backticks, which the page renders as code spans for a reader. The FAQPage structured data was serialising the raw string instead, so the copy an answer engine quotes verbatim contained the backtick characters. Structured data has no notion of inline code, so the markers are now stripped on the way into it.

• New surfaceBackticks stripped from FAQPage structured data on every page that emits it

  The site writes inline code in editorial prose with backticks, and InlineCodeText turns them into <code> spans for a reader. buildFAQSchema passed the raw string straight into the FAQPage JSON-LD, so the Cursor page told Google and every AI crawler to install Skills by dropping the folder under "`.cursor/skills/`" — backticks included — and five other answers had the same problem. Since the whole point of that structured data is being quoted verbatim by an answer engine, the markers are now stripped as the prose is serialised. One helper in the shared schema builder, so the fix lands on the agent, category, home, compare and matrix pages at once. Unpaired backticks are deliberately left alone, matching the renderer, so a stray one never eats the rest of a sentence. Nothing changes for a reader: the same answers still render as code spans.

  /skills-for-cursor →
• VerifiedA regression report withdrawn before it was made

  This started as an apparent contradiction: a check said the Cursor page still printed literal backticks, which the previous release had claimed to eliminate. It did not. The check stripped HTML tags but not script contents, so it was counting the page’s own JSON-LD and hydration payload as visible text — and an earlier run of it returned zero only because that fetch had been truncated. The visible page was correct all along and the previous claim held. What the wrong check did surface, once chased properly, was the real and much narrower bug above.

Every agent landing page opened by defining the agent — "AI-first IDE. SKILL.md works via the Cursor plugin layer" — which answers a question nobody asks. All seven now open with the picks for that agent and the constraint that actually decides them, and a build-time check parses the Skill names straight out of that prose and fails if any is not verified on the agent it is being recommended for. Fixing the leads also exposed that FAQ answers across eight surfaces, and the agent intros, were printing raw backtick characters to readers; both now render them as code.

• New surfaceAll seven agent pages lead with the picks, and a check stops them claiming false compatibility

  The pages rendered the agent’s one-line tagline as their opening paragraph, so a page targeting "best skills for Cursor" opened by explaining what Cursor is. Each now names its picks and the thing that decides them: Claude.ai leads with the document Skills and states that its sandbox cannot reach your machine, repository or shell; Claude API says outright that it is the thinnest surface by verified count because most Skills assume an interactive agent with a shell; Gemini CLI says its page is short on purpose, because few publishers document a Gemini install path, not because the CLI is weak; Antigravity says to check its pre-installed bundle before adding anything, since a second copy of a Skill is a second description competing for the same trigger. Because those leads name specific Skills, a new assertAgentLeadIntegrity() parses every backticked name out of the prose and fails the build if a catalogued Skill is recommended for an agent its compatibility[] does not list — the one class of error that would turn this improvement into exactly the false compatibility claim the site exists not to make. Verified by naming a Skill on the Gemini page that is not verified there and watching the build reject it. The taglines are kept, still feeding metadata and OG.

  /skills-for-cursor →
• New surfaceFAQ answers across eight surfaces were printing raw backticks

  The category intros had this fixed a day earlier, but the same pattern was still live in FAQ answers and in the agent-page intros: the prose writes file paths and Skill names in backticks, and the renderer passed the string straight into a paragraph, so readers saw the backtick characters. FAQ answers go through one shared component used by the category, agent, compare, matrix, skill, agent-guide and explainer surfaces, so the fix lands on all of them at once. Verified across thirteen representative pages spanning every one of those surfaces: zero literal backticks in visible text.

Added eleven verified skills. Education went from seven entries to eleven, finance from nine to thirteen, and sales from eight to twelve — which leaves healthcare, at five, as the only vertical in the taxonomy below ten. Two of the additions close gaps this site had publicly declared or quietly carried: norman-finance makes the monthly-reports guide’s "no bookkeeping Skill clears our bar" sentence false, so that sentence was rewritten rather than left to rot, and the sales vertical had CRM operation but nothing that finds a company or verifies a contact. Rejected three candidates on evidence: Pigment for a revocable licence, Coursera for having no licence at all, and a 396-star 244-skill GTM repo for being 244 slide outlines. Two payments skills were tagged to a vertical that does not exist, so a category slug with no page now fails the build.

• Added4 education skills — the vertical went from 7 to 11

  education-agent-skills is a 165-skill teaching set covering lesson design, assessment, differentiation and feedback; it was checked for padding before it was accepted — all 165 descriptions parse, none is an exact duplicate of another, every one carries an explicit "Use when" trigger, and the bodies run 7.7KB to 72KB — but it is CC BY-SA 4.0, so the ShareAlike obligation is stated in the entry rather than buried. research-workspace and zotero-skills cover the literature side (search, screening, extraction, and a real Zotero library), and research-co-pilot handles the systematic-review workflow. Every one is community-tier and that is what the badge says: no education vendor ships a first-party skill yet.

  /skill/education-agent-skills →
• Added3 finance skills — the vertical went from 9 to 13

  norman-finance is 19 MIT-licensed skills for the bookkeeping loop rather than the checkout: categorising bank transactions, matching receipts, monthly reconciliation, invoicing, chasing overdue payments, and DATEV and VAT preparation. It is Germany- and EU-specific, and the entry says so, because a US bookkeeping workflow gets little from the tax and filing skills. netsuite-finance-analyst is Oracle NetSuite’s own analyst skill from Anthropic’s official plugin marketplace, and carta-cap-table is Carta’s Apache-2.0 cap-table skill. The thirteenth entry is lemonsqueezy-claude-skills, which was already in the catalog and is now tagged to the vertical it belongs in.

  /skill/norman-finance →
• Added4 sales skills — the vertical went from 8 to 12

  The gap was the top of the funnel: the category had HubSpot for operating a CRM, a sales-team pack and Anthropic’s own sales plugin, but nothing that finds a company, finds the person, or verifies the address. vibe-prospecting is Explorium’s B2B data skill, confirmed by reading Anthropic’s own marketplace manifest rather than the vendor’s README — it is listed there, sha-pinned, under the name vibe-prospecting. gtm-skills is LeadMagic’s 211-skill library across 25 categories, where each skill ships a named framework, an output template and a deliverable checker; its provider comparison tables put LeadMagic first, and the entry says to read those rankings as a vendor’s. aigtm is 64 skills for the seller’s day — meeting prep, deal strategy, post-call summaries, pipeline health — with the strongest trigger discipline of anything in this sweep and an explicit instruction not to fabricate when a research step comes back empty. gtm-eng-skills is Deepline’s 16 skills for waterfall enrichment, chained across providers with cost-per-hit ordering and identity validation.

  /skill/vibe-prospecting →
• VerifiedA 396-star, 244-skill GTM repo rejected — count is not the bar

  gtmagents/gtm-agents has more stars than every skill added in this sweep combined and 244 SKILL.md files, and it was still rejected. The median skill body is 1.3 kB of bullet-point framework with no procedure and no bundled scripts; only 77 of its 244 descriptions carry a usage trigger, so most of them will never fire; three skill names are duplicated; and the last commit was 2026-04-03 with 98 issues open. Compare the entries that were accepted in the same pass: median bodies of 9.0 kB and 4.9 kB, with 168 of 211 and 64 of 64 descriptions carrying triggers. This is what "catalog count is not a metric" means when it costs something.

• VerifiedStar counts and install verification backfilled onto last week’s additions

  The previous batch shipped with no githubStars and no verified dates, on the stated grounds that the GitHub API was unreachable here and that every install path was a plugin-marketplace command. Both were wrong. Star counts came back first-hand from the API for six of the seven entries — the seventh is skipped deliberately, because netsuite-finance-analyst lives in a branch subdirectory of Oracle’s SuiteCloud SDK and that repo’s 335 stars belong to the SDK, not the skill. And the skills CLI installs from an arbitrary GitHub repo, so education-agent-skills, research-co-pilot and zotero-skills now publish a one-command install that was executed here today and carry a real verified date. The rest still carry none: where a skill’s engine is an MCP server or a vendor CLI, the working install is the plugin path, and that cannot be run in this environment.

• VerifiedA broken skill found by running the install — research-co-pilot ships 14 of 15

  Installing research-co-pilot surfaced a defect no amount of reading the README would have found: peer-review’s frontmatter puts an unquoted colon inside its description ("returns a reviewed file: .docx"), which makes the YAML invalid, so the installer skips that skill and the user silently gets fourteen of fifteen. The entry now says so in its not-ideal-for, and its install snippet names the skipped skill. This is the argument for executing install commands rather than transcribing them.

  /skill/research-co-pilot →
• New surfaceThe bookkeeping gap the site had declared is now described accurately

  The monthly-reports guide stated in its limits and again in its FAQ that no Skill in the catalog cleared our bar as a dedicated bookkeeping tool. Adding norman-finance made that false, so both sentences were rewritten in the same change: a bookkeeping set now exists, but it installs from a plugin marketplace in a terminal and is EU-specific, which is why it still does not appear in a guide whose whole premise is the no-terminal path.

  /guides/best-skills-for-business-reports →
• New surfaceTwo editorial decisions made and published: healthcare stays at five, and verticals mean “useful in this domain”

  Healthcare is the only vertical under ten and will stay there. The alternative was padding it with medical skill collections that did not clear the bar — the largest candidate, a 605-skill set, was rejected because one file is 99.7% identical to Anthropic’s with no attribution under a blanket in-house authorship claim — or de-listing a page whose top two answers are Anthropic’s own healthcare and life-sciences skills. Neither improves the answer a reader gets, so the page keeps its five entries and continues to say in its first paragraph that five is the thinnest vertical in the catalog. Separately, the verticals now state their tagging rule: a vertical lists the Skills useful in that domain, including general-purpose tooling commonly used there, not only Skills built for it. That is why a document parser appears on the finance and education pages and a text-to-speech skill on education. The narrow reading was considered and rejected because it would have dropped education from 11 to 6 for sharper but thinner pages; the site’s answer to a general-purpose skill outranking domain-native ones is to say so in the lead, which every vertical page does. The rule is rendered on every vertical page from one constant, so nine pages cannot drift into nine explanations.

  /top-skills-for-healthcare →
• New surfaceTwo head-to-head pages for the sales additions, and a link label that misdirected

  sales-plugin-vs-gtm-skills answers the question the sales vertical now poses: Anthropic’s nine official skills cover the motion an AE runs daily, LeadMagic’s 211 cover the motions a team runs quarterly — and one of them is published by a data vendor that ranks itself first in its own provider tables, which the page says outright. vibe-prospecting-vs-gtm-eng-skills is the buy-versus-orchestrate call in B2B data: one vendor’s coverage against a waterfall chained across 28 providers, with the crossover named — single-provider coverage runs 60-75% of contacts, chained runs 85-92%, so the waterfall stops being over-engineering at the volume where that gap costs you a campaign. Writing them surfaced a link that had been misdirecting: firecrawl-vs-browser-use offered to "browse the full Web Scraping category" and linked to Browser & Web Automation. A new build-time check now fails on any comparison naming a skill the catalog does not have, comparing a skill with itself, or labelling a category link with a name other than the category it points at.

  /compare-skills/sales-plugin-vs-gtm-skills →
• VerifiedA fix this changelog already claimed was only two-thirds done

  The 2026-08-26 entry below says the claude.ai install snippet was corrected to the documented flow, "replacing an outdated Settings → Features reference". The snippets were corrected; the prose was not. Seven more places still told readers to upload skills under Settings → Features — the Claude.ai agent tagline, two strings in its agent guide, two in its long-form intro, one install snippet, and the per-agent table in the how-to-install-skills guide. All seven now describe the real flow: enable code execution under Settings → Capabilities, then upload the zip under Customize → Skills. The earlier entry is left as written, because a changelog that edits its own history is worth nothing; this is the correction, forward. lib/claudeAiInstall.ts now lists every place the flow is described in prose, so the next label change updates all of them at once.

  /skills-for-claude-ai →
• New surfaceTwo collections were the same page; a third claimed the wrong licence

  best-open-source-skills filtered `open-source + builder-favorite`, and since every builder-favorite in the catalog is also open-source, that conjunction resolved to exactly the same seven skills as best-builder-favorites — two URLs, identical content, for as long as both existed. best-one-command-skills had the same bug in milder form, cut from 120 skills to 5. Both now filter on the axis they are named for. A new build-time check throws if any two collections ever resolve to the same set again, verified by reintroducing the bad filter and watching the build fail. Re-scoping the open-source page also exposed a false licence claim it had been hiding: remotion-best-practices is licensed Proprietary and flagged isOpenSource: false, yet carried the open-source filter, so it appeared on a page titled Best Open-Source Skills. Its filter is corrected, and the integrity check now also fails the build on any skill whose licence filter contradicts its own licence field.

  /collections/best-open-source-skills →
• New surfaceAll twelve collection pages now open with a recommendation, and their counts are derived

  A collection page was a one-line subtitle above a ranked grid, with no citable answer for the query it targets — "best open source claude skills", "most installed", "best skills for startups". All twelve now lead with the picks and the caveat that matters for that axis: the open-source page names the four licences held by a single skill each, because AGPL-3.0 and CC BY-SA 4.0 are the ones that change what you may ship; the action-taking page is framed as a risk register, since every entry there can change state outside your machine. Every number quoted in those leads is derived from the catalog at render rather than typed — the first draft was written from a hand-count that was silently wrong, because the script producing it could not resolve the shared-filter spreads in the Anthropic pack file and undercounted open-source as 148 instead of 192.

  /collections →
• New surfaceAll nine vertical pages now open with a recommendation instead of a definition

  The primary category pages have always led with a citable answer — "the best agent skills for document processing in 2026 are Anthropic’s docx, pdf, pptx and xlsx" — because an answer engine quoting two sentences needs those two sentences to contain the pick. Every one of the nine vertical pages instead opened by defining the category ("Sales vertical Skills handle outbound personalization…"), which is unquotable. All nine were rewritten to name the actual picks and, where a general-purpose Skill outranks the domain-native ones, to say so in the lead rather than let the page contradict itself: the education page now states plainly that the document and voice tools above its teaching Skills are general-purpose Skills useful in a classroom, not education Skills. Also fixed the rendering underneath: those intros have always written skill names in backticks and the page printed the backtick characters, on all twenty-nine category pages.

  /top-skills-for-sales →
• New surfaceA category slug with no page now fails the build

  stripe-directory and lemonsqueezy-claude-skills were both tagged to a "business" vertical that has never existed — the field is typed as string[], so the typo compiled cleanly and quietly dropped both skills out of every vertical page. lemonsqueezy is now in finance, where its refunds, receipts and revenue reporting belong; stripe-directory carries no vertical, because agent-initiated vendor discovery genuinely fits none of the nine. A new assertTaxonomyIntegrity() check runs at build time and throws on any category slug that resolves to no published page, and on any skill with no primary category at all. It was verified by reintroducing the bad slug and watching the build fail.

• VerifiedTwo candidates rejected on licence grounds — Pigment and Coursera

  Pigment ships planning skills under its own "Pigment Skills License v1", which we read in full rather than summarising from the README: the grant is revocable, it terminates automatically when you stop using Pigment’s Services and requires you to delete all copies, it forbids using the skills to benchmark or build a competing product, and it takes a perpetual irrevocable licence to any feedback you give. That is not a licence a directory can recommend installing. Coursera’s learning skill is listed in Anthropic’s own official plugin marketplace and has no LICENSE file of any kind, which under default copyright means no permission to use it. Neither is a quality judgement on the skills themselves, and both would be reconsidered if the terms changed.

• Added3 comparison pages for choices the catalog now forces

  hyperframes-vs-remotion-best-practices turns on authoring language and licence — Apache-2.0 HTML-to-MP4 against a skill for a framework whose own licence is proprietary above a revenue threshold. chrome-devtools-vs-playwright-cli is diagnosis against automation: both drive a real browser, only one has a test runner. liteparse-vs-docling is lookup against conversion, which is the actual fork in document extraction.

  /compare-skills/chrome-devtools-vs-playwright-cli →

Added 26 verified skills — the sales vertical doubled to eight, healthcare grew from two to five, and seo-content, web-hosting-deploys, payments-billing and email-messaging each reached ten, and a second sweep closed the last four nine-entry categories so nothing in the taxonomy now holds fewer than ten. Opened a second audience alongside developers: four guides written in the language a founder searches, a Business Owners collection of all 31 Skills that install without a terminal, and that install path published on every page where it applies. Catalog figures quoted in guide prose are now derived from the catalog instead of typed by hand.

• Added15 verified skills across sales, healthcare, SEO, hosting, payments, and email

  Sales: Anthropic’s official sales plugin (research → outreach → pipeline), HubSpot’s first-party CRM skills, the 1,074-star ai-sales-team pack, and a 37-skill HubSpot admin toolkit. Healthcare: Anthropic’s healthcare plugin (FHIR, ICD-10, prior-auth) and life-sciences marketplace, plus PhenoML’s standalone FHIR skill. SEO: SE Ranking’s 26-skill official suite and the 870-star Agentic SEO audit machine. Hosting: Railway and Render, both in Anthropic’s official plugin marketplace. Payments: PayPal’s AI toolkit and Dodo Payments’ 17-skill MoR suite. Email: Postmark and Loops, both first-party. Every repo was cloned, its SKILL.md frontmatter parsed, its licence file read, and its last-commit date taken from git on 2026-08-26.

  /skill/sales-plugin →
• Added11 more skills — the last four categories that stood at nine

  document-processing reached thirteen with LlamaIndex’s liteparse (local extraction, no API key), Box’s own skills including its contract-review workflows, and Carbone’s template engine for data-driven document generation. browser-automation reached twelve with the Chrome DevTools team’s own skill set, Bright Data for sites that fight scrapers, and Apify for running or packaging Actors. video-media reached twelve with HeyGen’s HyperFrames (HTML to deterministic MP4), Runway for generative video, and Canva for brand-checked output at volume. diagrams-visualization reached eleven with Mapbox’s data-visualization patterns — the category had nine entries and none could put data on a map — and the Tufte chart skill, a community entry ranked on the strength of a named design canon rather than a vendor SDK. No category now holds fewer than ten entries.

  /skill/chrome-devtools →
• VerifiedTwo licence re-checks flipped; two did not

  HeyGen’s hyperframes and Apify’s Claude plugin were both rejected in earlier sweeps for shipping no licence. Both now carry real LICENSE files — Apache-2.0 — and are listed. The Apify entry deliberately points at apify-claude-code-plugin rather than the apify/agent-skills mirror, which ships the same five skills with no licence at all. Figma and Cloudinary were re-checked and still fail: their plugin manifests name a licence but no LICENSE file exists in either repo, and a manifest field is not a grant. Between them they hold roughly twenty-nine on-topic skills that stay unlistable until that changes. Also rejected: Zyte (proprietary EULA), iFLYTEK (Chinese-only descriptions, re-hosted third-party skills), and every scraped aggregator catalog.

• VerifiedVerified dates only where the install command actually ran

  Seven of the fifteen carry a Verified date because their exact published install command was executed on 2026-08-26 (the npx and git-clone paths). The eight plugin-marketplace entries ship with repo, path, licence, and upstream date confirmed but no Verified date — same standard as the 2026-08-19 release — until their /plugin installs are run.

• VerifiedCandidates rejected, and why

  ColdIQ’s 90-skill GTM pack (264 stars) has no licence file and gates key skills on a paid API. Chargebee’s integration skill still ships without a licence — watch-listed, not listed. Salesforce’s sf-skills and Pipedrive’s app-scaffolding are developer tooling, not sales work. OpenClaw’s 869-skill medical aggregation is the catalog-over-curation play this site exists to counter. AIPOCH’s medical set came off the watch-list as a rejection, and the check is worth recording precisely: its README states that every skill is “originally designed and developed in-house”, but its clinical-trial-protocol-skill is 99.7% identical to Anthropic’s, re-badged with author: AIPOCH and a licence header, pointing at AIPOCH as the source, with no attribution anywhere in the file. To be fair to them, that was the only near-duplicate: all 605 of their skills were compared against Anthropic’s life-sciences repo and K-Dense’s 163-skill collection, and the other 604 — including 58 that share a name with a K-Dense skill — are independently written at 3–8% similarity. One re-badged file is still enough: a blanket authorship claim this site would repeat has to be accurate, and a 605-skill repo is the catalog-over-curation shape declined above. Two accepted entries carry honest staleness notes: ai-sales-team (last push March 2026) and postmark-skills (February 2026).

• Verifiedscientific-skills repo rename fixed

  K-Dense renamed claude-scientific-skills to scientific-agent-skills; the old URL browser-redirects but API and CLI lookups against it fail. The entry now points at the new repo (163 skills, MIT, pushed 2026-08-24) and its full install was re-executed today.

  /skill/scientific-skills →
• New surfaceA no-terminal install path on every claude.ai-compatible skill

  The claude.ai install snippet now describes the documented flow — Customize → Skills → upload the skill folder as a zip, with code execution enabled under Settings → Capabilities — replacing an outdated Settings → Features reference. The works-with-claude-ai filter was backfilled from 14 to all 31 skills whose verified compatibility includes claude.ai, and a new Business Owners collection lists all 31 with the failure mode stated: server-side sandbox, no shell, no local filesystem.

  /collections/best-skills-for-business-owners →
• New surfaceFour guides in the language business owners actually search

  A separate guide cluster for people who are not developers: best Skills for writing client proposals, for your newsletter, for social media content, and for building a pitch deck. Every pick on these pages carries verified Claude.ai compatibility, enforced at build time rather than by review — a pick that needs a terminal fails the build instead of shipping under a no-terminal promise. Each page states the limits too: these Skills draft, they do not send or post, and the sandbox they run in has no shell and no access to your files.

  /guides/best-skills-for-client-proposals →
• New surfaceThe browser install, published on the skill pages themselves

  Skill pages whose verified compatibility includes Claude.ai now carry a collapsed four-step browser install — enable code execution, download the folder, upload the zip under Customize → Skills, then describe the job — so a reader arriving from an AI answer is not handed a command line they will not use. The same four steps are published as HowTo structured data on the business-owner guides, and they live in one module rather than being retyped per page.

• AddedDevOps × Codex matrix page opened

  The combination had cleared the six-skill gate but had no hand-written lead, so it had never shipped. It says something the category page cannot: what is verified on Codex here is the operational half — GitGuardian secret scanning, Sentry error handling, deploy-checklist, incident-response, Redis caching, Firebase hosting — while the Terraform and Kubernetes skills that anchor the same category on Cursor are not verified there. The FAQ states that absence is policy, not failure.

  /top-devops-infra-skills-for-codex →
• VerifiedCatalog counts in guide prose are now derived, not typed

  Fifteen sentences across the per-agent guides claimed "134 catalogued Skills" and figures like "77 of 134" — numbers that were already wrong before this release and got worse with every addition. A number nobody updates is worse than no number, because the reader cannot tell it is stale. All of them now render from the catalog itself, so Cursor reads 78 of 180 today and reads the truth tomorrow without anyone remembering to edit prose.

• New surfaceMobile funnel rebuilt

  Phones now get a navigation drawer and a persistent search entry, one-tap copy on every install command, a sticky install bar on skill pages, a web app manifest with real icons, and 49 kB gzip less JavaScript on every skill page after a barrel-import fix. Four guide tables that clipped install commands at phone widths now scroll.

Closed every category that held fewer than eight entries by adding 31 verified vendor-published skills, replaced build-time sitemap dates with dates derived from the catalog, published a per-agent guide series and six explainer guides, opened the gated category-by-agent matrix, and made the catalog available as a single JSON dataset with an embeddable ranking badge.

• Added31 verified skills — every thin category closed

  Eight categories held fewer than eight entries. All are now at nine or more, filled almost entirely with vendor-published skills verified against primary sources: Semgrep, GitGuardian, OWASP and Snyk in security; TigerData Postgres, Qdrant, Weaviate, Turso and Bytebase in databases; Microsoft Playwright and Browserbase in browser automation; Slack and Knock in messaging; two more Stripe skills including agentic checkout; draw.io and Microsoft in diagrams; Docling, Microsoft Azure and markdown-exporter in documents; ElevenLabs speech, transcription and dubbing in media.

• VerifiedEleven candidates rejected

  Recorded so the gaps are explainable. OpenAI's 19-skill data-visualization suite is the richest diagram content in the ecosystem but its host repo is archived and carries no licence file. Figma's diagram skill, HeyGen's video skills, Apify, Polar and Chargebee all failed the licence check — a README badge or a frontmatter field is not a licence grant, and no licence file means all rights reserved. Google Workspace's 95 auto-generated CLI skills have five-word descriptions that cannot trigger reliably. Chainguard has no documented install path; Endor requires a commercial tenant.

• ScoringVerification standard stated precisely

  The methodology page now separates the two dates on every entry and states that a Verified date is set only once the install command has been executed. The 31 skills added in this release ship with the repo, path, licence and upstream date confirmed but no Verified date, so they carry the verification-score penalty until their installs are run. An empty field is worth more than a badge that means "probably fine".

• New surfaceHonest lastmod everywhere

  The sitemap previously stamped the build date on every non-skill URL, which teaches search engines to discount the signal — including the per-skill dates that were real. Every lastmod is now derived from the content behind the URL: category pages from the newest change among their skills, guides from their own edit date, editorial pages from a dated registry.

• New surfaceGuide dates unfrozen

  All 22 best-for and deploy guides shared one hardcoded publish date in the article schema. Each now carries its real first-publish date and its real last-content-change date, both shown on the page.

• New surfacePer-agent guide series

  Seven new guides answering "best skills for <agent>" — Claude Code, Cursor, Codex CLI, Claude.ai, Claude API, Gemini CLI, Antigravity — each with an install model, ranked picks, and an explicit limits section naming what is not verified there.

  /guides/best-skills-for-cursor →
• New surfaceSix explainer guides

  Skills vs Cursor Rules, Skills vs AGENTS.md, marketplaces compared, how to audit a Skill before installing, whether Skills cost tokens, and how to publish and get listed.

  /guides/skills-vs-cursor-rules →
• New surfaceCategory-by-agent matrix

  Fifteen pages answering "top <category> skills for <agent>" for Cursor and Codex, behind a quality gate: at least six verified-compatible skills, a hand-written lead, and an independent third-party agent. Claude Code is excluded because every catalogued skill is verified there, so the page would duplicate the category.

• New surfaceOpen dataset and ranking badges

  The full catalog is now available as a single machine-readable file at /skills.json, and every skill page offers an embeddable SVG ranking badge that publishers can drop into a README.

  /skills.json →
• New surfaceSkill-vs-skill comparisons: 3 to 9

  Six new head-to-head pages, each on a rivalry the new entries created: docling vs pdf (reading versus writing documents), semgrep vs trail-of-bits (findings versus judgement), Playwright CLI vs Browserbase (local versus hosted browsers), postgres vs supabase-skill (engine versus platform layer), drawio vs mermaid (editable artefact versus text in the repo), and Stripe Connect vs Paddle (owning tax compliance versus renting it). Comparison coverage went from 3 categories to 6.

  /compare-skills →
• VerifiedOpenClaw and ClawHub documented, not yet catalogued

  Verified from primary sources: OpenClaw reads SKILL.md natively with the same two required frontmatter fields, loads skills from <workspace>/skills, <workspace>/.agents/skills and ~/.agents/skills in that precedence order, and has a public registry (ClawHub) installable as openclaw skills install <package> or npm i -g clawhub. It is covered in the install guide and the marketplaces comparison. It is deliberately NOT added as an eighth agent target: per-agent compatibility here means the skill was tested there, no such pass has been run, and OpenClaw’s own docs decline to claim Claude Code skills work unchanged — they note Codex’s skills directory is not an OpenClaw root and ship migration commands instead. An agent page listing zero verified skills would be the hollow page this site removes categories for.

  /guides/how-to-install-skills →
• New surfaceChangelog and Atom feed

  This page, plus a feed at /feed.xml so changes are discoverable without polling the sitemap.

  /feed.xml →

A verification sweep against primary sources removed eleven entries whose install commands pointed at organisations that do not exist, and corrected the repository and provenance data on several others.

• RemovedEleven community entries

  cro-patterns, email-sequence-builder, docker-best-practices, github-actions-builder, aws-cdk-patterns, cypress-testing, vitest-skill, typescript-strict-mode, tailwind-best-practices, standup-skill, and mindmap-generator were removed: each carried a <topic>-skills/ style install command pointing at a GitHub organisation that does not exist, and no credible canonical repository could be found. Each URL now permanently redirects to its nearest live category.

• Removedweb-scraping-patterns, puppeteer-patterns

  Removed for fabricated publisher organisations that could not be verified against any real repository. Both redirect to Browser & Web Automation.

  /top-skills-for-browser-automation →
• Removedget-shit-done

  Removed because its upstream repository is archived. Redirects to Workflow Orchestration.

  /top-skills-for-workflow-orchestration →
• VerifiedRepository and provenance corrections

  Real repositories wired for several verified-org entries (Shannon, shadcn, Sentry), and provenance tiers corrected where the published claim was stronger than the evidence supported.

• AddedDuckDB and Neon

  Two official vendor database skills added to deepen Databases & Schemas.

  /top-skills-for-databases-schemas →

The Domains & DNS category was removed rather than filled: no SKILL.md-format domain or DNS skill cleared the credibility bar, and a hollow category page is worse than no page.

• RemovedDomains & DNS category

  Removed with a permanent redirect to Web Hosting & Deploys. The alternative — filling it with one or two weak entries — would have pulled down the average and misrepresented the state of the ecosystem.

  /top-skills-for-web-hosting-deploys →

Added the official Paddle and Lemon Squeezy skills so the payments category covers merchant-of-record billing, not only Stripe.

• AddedPaddle and Lemon Squeezy

  Vendor-published skills covering merchant-of-record billing, tax handling, and live store operations.

  /top-skills-for-payments-billing →

Hookdeck and Twilio skills verified against their repositories and added, with Composio for broad third-party routing.

• AddedHookdeck, Twilio, Composio

  Webhook receiving, messaging, and multi-service tool routing — each verified against the publishing organisation's own repository.

  /top-skills-for-apis-integrations →

Four verified deploy skills added and a six-page deploy cluster published covering Vercel, Netlify, Cloudflare, and GitHub Pages.

• AddedFour deploy skills

  First-party deploy skills for the major hosts, each with a tested install command.

  /top-skills-for-web-hosting-deploys →
• New surfaceDeploy guide cluster

  A pillar guide plus five per-host guides for shipping a site from chat.

  /guides/deploy-a-website-with-ai-agent-skills →

Four Resend-published skills added, covering transactional sending, deliverability, and React email templates.

• AddedFour Resend skills

  Vendor-published and verified — transactional email, deliverability practice, React Email templates, and agent inbox handling.

  /top-skills-for-email-messaging →

Four Stripe-published skills added, and the SEO & Content category deepened.

• AddedFour Stripe skills

  Integration best practices, payments implementation, the skill directory, and the API-version upgrade guide.

  /top-skills-for-payments-billing →

The Anthropic plugin-pack and built-in expansion added 41 official Skills across design, marketing, engineering, and productivity.

• AddedAnthropic pack expansion

  Design, marketing, engineering, and productivity Skills published by Anthropic, catalogued separately from the hand-curated base set so provenance stays legible.

  /collections/best-official-skills →

The directory launched with a hand-curated catalog, a published scoring rubric, per-agent install snippets, and agent-readable exports.

• ScoringRubric version 1.0

  Nine weighted inputs, published rather than implied: install count, provenance, stars, recency, verified cross-agent compatibility, documentation depth, install ergonomics, licence, and verification freshness.

  /about/methodology →
• New surfaceAgent-readable exports

  llms.txt, llms-full.txt, and per-skill .md and .json variants from day one.

  /llms.txt →