Source: raw/8_Claude_Skills_I_Can_t_Live_Without.md, raw/12_Claude_Plugins_Skills_MCP_s_I_Can_t_Live_Without.md

Creator: Ben AI · Platform: YouTube

Two roundup videos from Ben, an AI agency owner who works mostly in Cowork for knowledge work and switches to the Code tab for CLIs. The first covers eight “meta skills” that improve Claude on any task, not one workflow. The second covers twelve tools that extend what Claude can reach. Most items already have their own wiki article, linked below; this page records Ben’s specific use for each and the usage tips that are new. He gives the skills and install links away as a free bundle and sells a larger library in his paid community.

Key Takeaways

  • Chain the quality skills onto your writing skills. Ben adds a last step to his LinkedIn and newsletter skills that runs humanizer (and optionally fact-checker) before showing the output: “Can you update my LinkedIn post skill to always run the humanizer skill after it’s generated the initial output?”
  • Make prompt optimization automatic. Put a rule in Cowork’s global instructions or a folder CLAUDE.md: “Every time I use a medium-to-long unstructured prompt for a task, auto-run the prompt master skill on the input prompt.” In his demo Claude applied it without being asked.
  • Prefer a CLI to an MCP when one exists, for two reasons Ben gives: fewer limits (he says the Gmail connector was read-only and could not create Sheets, Docs or Calendar events, while the Google Workspace CLI can) and lower token cost. CLIs work best in the Code tab.
  • Use Caveman on input, not output. Compress files Claude reads every session (CLAUDE.md, frequently used skills, large context docs, MCP documentation) and test the compressed version against the original. Ben reports a 40% shorter CLAUDE.md and 18% fewer skill input tokens “with no meaningful quality regression.”
  • For browser work you repeat, switch from Claude in Chrome to a script-based tool (Playwright CLI) or a generated CLI (Printing Press). Ben calls them faster, less error-prone and cheaper.

The eight meta skills (April 2026)

SkillWhat Ben uses it forWiki coverage
Process interviewer (Ben’s own)Plans a skill or complex task by asking 10 to 15+ questions; checks the second brain first so it skips answered questions; ends with a skill brief (PRD) that includes Anthropic’s best practices. Adapted from Matt Pocock’s grill-me. Ben budgets 30 to 45 minutes and answers by voice.Skill-Engineering Framework, grill-me
Prompt master (found online)Turns a messy voice-dictated brain dump into a structured prompt; with a second brain, pulls in voice and business context.Not covered elsewhere; repo not named
HumanizerRuns on anything other people will read: emails, community guides, LinkedIn posts, video intros. Built on Wikipedia’s “Signs of AI writing” guide. Add your own banned words over time.Nine Free Agent Skills § Humanizer
Fact-checkerThree uses: check AI output (verifies each claim against the source transcript and the web, reports true / mostly true / unverifiable / false), check his own video outlines before filming (it caught two errors in a Claude Routines outline in Notion), and check other people’s articles.Anti-AI Slop Guide § Beyond UI
Find skillsSearches the skills.sh ecosystem with install counts, then installs the pick. Ben cites “more than 91,000 skills.”skills.sh
Frontend slidesAnimated HTML decks, or converting PPT to HTML. Without a brand guide it picks a preset (on AI topics, almost always Neon Cyber); attach your brand guide or customize the skill once.Frontend Slides
Decision toolkitA “guide, don’t decide” framework for big decisions; produces a markdown file and an interactive HTML wizard (frame the decision, start-fresh test, timing, who is affected, bias checks, opportunity cost, scenarios, first-principles test). His example: whether to raise his community’s price.Not covered elsewhere
MCP builder (Anthropic built-in)Builds an MCP server for software with an API but no connector. Ben built ones for Circle (post, thread, analyze issues) and Reddit. Add the generated JSON via Settings → Developer → Edit config.skills

Bonus skills in his bundle: an agent-browser skill he calls more efficient than Claude’s browser use, an audio transcriber for any web media player, and an OpenRouter skill for comparing outputs across model providers.

The twelve tools (May 2026)

ToolTypeBen’s use and tipsWiki coverage
Google Workspace CLICLIHis most-used. Example: “check my inbox and make a Google Sheet with all sponsorship opportunities from the last 90 days.” Setup needs a Google OAuth app; his free setup skill drives the browser through it.Higgsfield as a Creative Agency (GWS CLI as asset database)
HiggsfieldMCP (Cowork) + CLI (Code)Image (Nano Banana, GPT Image 2) and video (Seedance) inside Claude. His team makes community banners with a brand guide and one reference image; his thumbnail skill keeps his photo and logos in the skill. Ben says the $15 plan covers 100 Nano Banana Pro images or 23 videos.Higgsfield (Nate Herk)
Printing PressPlugin + CLI library50+ prebuilt CLIs, build your own from any MCP or software, including ones without an API. He generated a YouTube Studio CLI.Printing Press
ImpeccableSkills (Claude Code)Click any element of a local HTML page and run layout, adapt, polish, animate; “go” generates three variants. Add a DESIGN.md and “detect” fixes against it.Impeccable v3
VercelCLIDeploys an HTML file to a live URL in about 10 seconds. His agency uses it for client onboarding pages, dashboards and proposals; one skill reads a call transcript, builds a branded page and deploys it.skills.sh (Vercel)
CavemanPlugin (skills)Compression; see Key Takeaways. Claims “up to 75%”; Ben says his savings were lower.Token Optimization (measured savings)
FirecrawlMCP + CLIDefault scraper. In his test Claude’s web fetch failed on all five e-commerce sites; Firecrawl got four. 1,000 pages/month free. Cannot scrape social media.Firecrawl
ApifyMCPFor social media, which Firecrawl can’t scrape. Named as a bonus.Cowork + Apify recipe
Playwright CLICLIBrowser automation through scripts it writes and reuses. Demo: filled the contact forms of five local businesses for lead generation. Free, open source; an MCP also exists.Not covered elsewhere
Claude video pluginPluginDownloads a video and takes screenshots so Claude sees it, not only the transcript. Demo: a Google Sheet analysing AI creators’ Instagram Reels with the hook frame and why each worked.claude-video (likely the same tool^[inferred])
Vibe ProspectingConnectorLead database built for agents, filterable by intent signals (hiring, funding). Demo: five US SaaS companies, 10 to 50 staff, hiring BDRs now. Credit-priced; shows cost before each query.Cowork + Apify recipe § Vibe Prospecting
UnipileMCPRead and send on WhatsApp (unofficial API, QR-code login with your own number), Instagram and LinkedIn DMs. $55/month per Ben. He sent a test LinkedIn message to his co-founder. Watch LinkedIn’s daily limits (he believes 20 a day; he says to recheck).Mentioned in Shapiro architecture

Caveats

  • The MCP token claim is overstated for Claude Code. Ben says every chat loads full MCP tool information. Claude Code’s Tool Search loads MCP schemas on demand (CLI vs MCP). ^[ambiguous] The cost argument holds more strongly for Cowork connectors, which Ben was describing.
  • The Gmail connector claim is dated. In May 2026 Ben said the Gmail connector was read-only. By July, Eliot Prince’s recorded skill created Gmail drafts through the connector (Record a Skill), and in August Ben’s own triage skill used it to label, mark read and draft. ^[ambiguous] Check the connector’s current tool list before choosing a CLI for this reason alone.
  • Printing Press examples. Ben describes the library as CLIs for software “that has no public API yet” but names ElevenLabs, OpenRouter, DigitalOcean, eBay and Shopify, which do have public APIs. The library covers both kinds.
  • Caveman savings vary. The wiki records an independent test where Caveman cut about 8.5% of output tokens against an advertised 65% (Token Optimization). Ben’s 18% skill-input figure is a single run.
  • Unsolicited contact-form filling at scale may breach site terms or anti-spam rules; Ben shows five forms and doesn’t discuss this.

Try It

  1. Pick your two most-used writing skills and add a final step that runs humanizer, then fact-checker, before the output appears.
  2. Add the “auto-run prompt master on long unstructured prompts” rule to your global instructions (any prompt-improver skill works) and watch the next few long prompts.
  3. Run /context in a fresh chat, then compress CLAUDE.md with Caveman, A/B the old and new versions on one task, and keep the new one only if output holds.
  4. List any software you reach through Claude in Chrome more than weekly; try Playwright CLI or Printing Press on the most frequent.
  5. For a tool without a connector that has an API, ask Claude to “use the mcp-builder skill to build me an MCP server for ”.

Open Questions

  • Prompt master and decision toolkit: Ben does not name the repos; they come in his bundle.
  • Vibe Prospecting and Unipile pricing (credits; $55/month) and the LinkedIn 20-per-day limit are Ben’s figures, not checked.
  • skills.sh “91,000 skills” was not checked against the registry.