AI For Dev

summary-dubibubii-claude-skills-2026

4 min read

The Only Claude Skills You Need in 2026

Channel: Dubibubii Format: YouTube video URL: https://www.youtube.com/watch?v=2BFN2DtcQMw Length: ~17 minutes Published: 2026-04-08 Sponsor: none disclosed (creator promotes own Discord community)

Summary

Rapid-fire walkthrough of 33 Claude Code skills, MCP servers, and OSS repos the creator considers worth installing out of “500,000+” available in the broader skills marketplace. Mix of Anthropic-official skills, community skill packs, MCP servers, and standalone OSS projects. The source’s structural value is not its individual recommendations but the sheer scale of the plugin ecosystem it surfaces: the wiki has been tracking individual plugins one source at a time, and this is the first source to treat the plugin layer as a category in its own right. Anchors the new agent-plugins index page.

Key Points

  • Skills vs MCPs vs repos — the creator’s framing: “Skills improve Claude, MCPs connect Claude, repos host the tools.” This is the cleanest one-line distinction the wiki has captured.
  • Frontend Design skill is the most-installed Claude skill at 277,000 installs. Bans Inter / Roboto / Arial fonts and forces Claude to commit to a design direction before writing code.
  • Superpowers (superpowers) was accepted into the Anthropic marketplace in January 2026 and is now at 100K+ stars. Will delete code written before tests exist (not archive — delete).
  • Karpathy’s autoresearch hit 50K stars in a single week. Research-loop pattern: give it a goal, it runs experiments overnight.
  • Context7 (context7) is “the #1 MCP server by a mile” — 50K stars, 240K weekly NPM downloads. Creator’s universal recommendation: “if you install one MCP server from this entire video, make this the one.”
  • Garry Tan’s gstack (gstack) at 40K stars; TechCrunch covered it; “god mode” CTO quote went viral. 15 opinionated prompts structuring Claude into different roles (CEO, designer, EM, release manager).
  • Taskmaster (task-master-ai) exposes 36 MCP tools, ingests a PRD and breaks it into tasks-with-dependencies. Works in Cursor, Claude Code, Codex, “anything.”
  • Three Anthropic-official skills: PDF Processing, XLSX, PPTX. Plus the meta Skill Creator that generates skill files from plain-English descriptions.
  • Promptfoo is the wiki’s first dedicated prompt-security tool — red-teaming, edge-case testing, vulnerability scanning. Connects to ai-personal-agent-hardening as an offensive-tooling-for-defenders entry.
  • Container-Use by Dagger (Solomon Hykes / Docker creator) — agent sandboxing. Each agent gets a containerized environment that physically can’t break the host system. Direct support for the ai-personal-agent-hardening human-in-loop quarantine rule.
  • Claude Squad runs multiple Claude Code agents in parallel terminal sessions. Also supports Codex and Aider. 65K stars.
  • Firecrawl (firecrawl) at 97K stars (YC-backed) — passed Scrapy’s star count, the long-time default web scraper. “If your agent needs to read the internet, this is the standard.”
  • LangFlow at 146K stars — visual drag-and-drop AI agent pipeline builder, DataStax-backed. Sibling to n8n, dify, sim-ai in the visual-builder cluster.
  • Awesome Claude Skills (Travis Vian, 95K stars) and the official Anthropic skills repo (100K+ stars) are the canonical curated lists. Plus three marketplaces: SkillsMP (500K+ skills, largest catalog), SkillHub (30K AI-rated for quality), and the Maggie archive (daily fresh repo feed).
  • Install pattern: paste the GitHub link or skill URL into Claude Code; it figures out the rest.

Sponsorship & Bias Notes

Sponsor: none disclosed. The creator promotes their own Discord (“vibebuilders,” 1,800+ members) and a personal “vibe coding to $100K” challenge. The video is a creator-driven recommendation reel, not a paid placement.

Product placement / affiliations: none observed as paid, but the source is inherently editorial-opinion-driven. The creator self-identifies as a former marketer who “went through thousands” of skills and surveyed his community. Star counts and install numbers are claims-from-source, not independently verified by the wiki. Treat all numerical claims as approximate.

Comparison bias: the video does not compare competing plugins head-to-head; it presents each as a standalone recommendation. The lack of comparison means the wiki cannot file any “X beats Y” claims from this source.

Notable Quotes

“Skills improve Claude, MCPs connect Claude, repos host the tools.” — the cleanest distinction in the source

“It will literally delete your code if you write it before the tests exist. Not archive it, delete it.” — on superpowers

“If you install one MCP server from this entire video, make sure it is this one.” — on context7

Connected Pages

See Also

Graph View

Backlinks