Taste Skill 2.1 didn't load for a day. One colon did it.

THE SIGNAL / 01

Taste Skill 2.1 shipped this weekend. For about a day, Claude Code couldn’t load it, and nothing told you.

The cause was one colon. 2.1 rewrote the skill’s description, and the new text included “Also for app screens: dashboards…” without quotes around it. In YAML, a colon followed by a space inside an unquoted value breaks the frontmatter. Claude Code couldn’t parse the file, so it dropped the skill. It wasn’t in the skill list, calling it by name gave “Unknown skill”, and the only trace was a warning in claude --debug output. A user reported it at 00:16 UTC on October 19, and the maintainer fixed it in 2.1.2 five and a half hours later. The broken description had been on the main branch since 02:43 UTC on October 18.

If you installed or updated Taste Skill on Sunday, update it again, then check that it shows up in your skill list. If you write skills, quote your description or use a folded block (description: >-, which is what the fix did). The difference is two quote marks:

description: Also for app screens: dashboards… ← breaks

description: "Also for app screens: dashboards…" ← loads

The maintainer now runs a strict YAML parse of every SKILL.md on each push. That’s a test worth copying.

The release itself changes what you get when you install the repo:

1. One skill in the repo. The older styles (soft, minimalist, brutalist and the rest) moved to their own repo, Leonxlnx/taste-skill-classics, with the same install names. They run only when you call them by name: in Claude Code they’re set to disable-model-invocation: true, so the agent can’t reach for an old one by accident.

2. Redesigns moved into the main skill. Redesign Existing Projects is retired. Taste Skill now takes a before shot in a real browser, keeps your stack and content, and proves the change with an after shot.

3. It builds app screens now, not only pages: dashboards, settings, inboxes. The image-generation skills moved to their own repo and kept their install names.

• Read the changelog, then update. →
• View Taste Skill & platform support →

NEW & VETTED / 02

Five official servers, from the makers themselves, replace entries we removed.

Last week we said these would come back once we’d verified them. Every config below is copied from the maker’s own page, and each page’s auth line comes from the maker’s docs, not from us. We checked where each config comes from; we didn’t connect to or test the servers themselves.

• Airtablesearch bases, read and write records, manage automations. OAuth or token. What it can change follows your Airtable role.
• Hugging Facesearch models, datasets, Spaces and papers, and call Gradio Spaces as tools. OAuth or token. Its VS Code snippet is missing a closing brace, so we show the two Claude Code commands only.
• Replicatesearch, compare and run models through Replicate’s API. API key: you paste your token into a sign-in page on mcp.replicate.com, Replicate’s own domain. Its docs never call that OAuth, so we don’t either.
• falgenerate images, video and audio with fal’s models, and check their pricing first. OAuth, with no key to paste. The server is free; generations use your fal credits.
• Excalidrawhand-drawn diagrams that render inline in the chat, in clients that support MCP Apps. No key. Its README’s setup script changes into the wrong folder, so the page sends you to the README rather than copy it.

More are coming. Our re-check found official servers for about 30 more of the removed entries, and each one gets the same check before it’s listed.

THE PICK / 03

Claude API — Anthropic

S-RANK · ANTHROPICS/SKILLS 182,000+ STARS · 5 CONTRIBUTORS · UPDATED OCTOBER 9, 2026 BY CJ AVILLA

If you build on the Claude API, your agent’s memory of it is probably out of date. This skill exists to stop the agent from writing code from that memory.

It opens with a table titled “Your Training Prior May Be Stale”. Each row pairs a pattern an agent is likely to remember with what the API does now. For example, budget_tokens for extended thinking is now “rejected with a 400” on the newest models, which take thinking: {type: "adaptive"} instead. The rule under the table is blunt: “Never guess SDK usage.” Function names and import paths must come from the skill’s per-language files or the official SDK repos. That covers Python, TypeScript, Go, Java, Ruby, PHP, C# and cURL.

What made it the pick is the subcommands. /claude-api migrate moves your code to a newer model, and it asks which files are in scope before it edits anything. prompt-audit finds instructions in your prompts and CLAUDE.md that were written for older models. cost-optimize and hillclimb both work against your real bill and your evals, under a rule we’d like to see everywhere: “every run that exercises the model spends real money, so get the user’s approval first.”

The caveat: it is big and it is eager. Its SKILL.md alone is over 100 KB, and its description tells the agent to load it whenever a prompt names Claude in any form, or for any LLM-shaped task where you haven’t named a provider. It skips itself when your project uses another provider’s SDK. Its frontmatter pre-approves no tools.

$ npx skills add anthropics/skills@claude-api

• View skill & platform support →

DIDN’T PASS / 04

62 MCP servers left the directory on October 20. 17 we had marked for removal stayed, because we were wrong about them.

• The 62 publish no repository and no website.We announced this in Issue 004 and said it again in 006 and 007. If you installed one, your config keeps working, because it points at the server, not at us. Old links to their pages now go to their category, or to the official replacement where there is one (see 02). With the five new entries, the directory lists 139 MCP servers, down from 196.
• 17 did publish a source.We missed it. Re-checking each entry before the removal, we found that 17 had a public repository our entries didn’t link to, among them Desktop Commander, Repomix and Basic Memory. They stay, now with the link. XcodeBuildMCP has been renamed MobileBuildMCP and now lives under Sentry’s GitHub organization.
• We said October 14.The re-check was done on the 13th; we didn’t merge the removal until October 20.

WORKFLOW / 05

Marketing & Growth Engine — Corey Haines’s marketing skills chained into one loop: write the copy, fix the page it sits on, audit the SEO, plan what comes next, and measure it. Beginner.

Two of the five changed this week (see 06). Copywriting now reads a ledger of what you’ve learned about your audience before it drafts. Analytics Tracking got its own update. Read both commits before you update, the same as any skill.

The five skills, in the order the workflow runs them: copywriting, cro, seo-audit, content-strategy, analytics. To install just those five:

$ npx skills add coreyhaines31/marketingskills --skill copywriting cro seo-audit content-strategy analytics

• Marketing & Growth Engine →

WHAT CHANGED / 06

51 skills you may have installed were updated this week.

“Updated” means the skill’s own directory got a commit since October 12, not a judgment about what the commit did. Before you run update, read the commit. Seventeen are Microsoft’s Azure and Foundry skills, all in one commit on October 19. Six are Corey Haines’s marketing skills, changed on October 14 and 15. Top of the rest by stars:

• shadcn/uiOctober 14
• Agent SkillsOctober 13
• Taste SkillOctober 19 (see 01)
• ImpeccableOctober 18
• Last 30 DaysOctober 18
• HumanizerOctober 16
• Scientific SkillsOctober 19
• Planning with FilesOctober 17

…and 20 more. The date is on every skill’s page, under Last commit.