Claude Skills Part 2: 7 More Integrations Worth Installing

Part 1 of this series looked at five Claude Skills worth installing: Clay, the taste-skill, open_deep_research, OpenBB, and exo. That list leaned toward power users — people who already had Claude Code running and wanted to push it into research, market data, or distributed inference. The response from readers was clear: where's the rest? What about context budgets, memory, writing hygiene, and business ops?

This is Part 2. Seven more skills and integrations, verified against their actual repos. Some are official Anthropic plugins. Some are community builds that went viral. A few of the marketing claims you'll see on social media are overclaimed — I'll flag those. The install commands are exact. The star counts are as current as searches can get in June 2026.

---

Caveman — Compressed Output, Not Compressed Thinking

Token limits are real friction for anyone running long agentic loops in Claude Code. The JuliusBrussee/caveman skill addresses output token bloat directly: it instructs Claude to strip articles, pleasantries, hedging language, and filler phrasing while keeping every technical detail, code block, error string, and variable name intact. The result is responses written in a compressed, telegraphic style — "me fix bug, here code" territory — that the repo estimates cuts output tokens by 65–75% on benchmark tasks, with a stated range of 22–87% depending on the task type.

The mechanism is purely prompt-side. Caveman is a SKILL.md instruction set, not a tokenizer patch or a compression algorithm. It won't reduce your input context — it only shrinks what Claude writes back. The repo has hit ~70,000 GitHub stars and briefly trended #1 on GitHub and Hacker News. It ships with intensity levels (lite, full, ultra, wenyan variants), works across Claude Code, Cursor, Windsurf, Gemini CLI, and Codex, and takes about 30 seconds to install. For teams where Claude is generating reports, explanations, and code comments at volume, the savings on a Sonnet or Opus tier are meaningful.

Best for: Agentic loops, batch processing jobs, cost-sensitive Claude Code setups where output verbosity is the main drain.
Limitations: Input context is unchanged. Doesn't help if your bottleneck is a long codebase or large file injection. Caveman output can make logs harder to read for humans.
Cost: Free, open source.

Install:

curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash

Activate in Claude Code by typing /caveman. Deactivate with "normal mode."

---

claude-mem — Session Memory Without Manual Upkeep

Claude Code's native CLAUDE.md is a static context file — you write it, it loads, it doesn't update itself. thedotmack/claude-mem fills the gap by wiring five lifecycle hooks (SessionStart, UserPromptSubmit, PostToolUse, Stop, SessionEnd) into Claude Code so that what Claude does during a session gets compressed, indexed, and injected back into future sessions automatically.

The implementation runs a background worker service (requires Bun) that captures tool usage observations, runs an AI summarization pass, and stores structured memory. When a new session starts, relevant context is silently injected via hookSpecificOutput.additionalContext — you don't see a prompt, Claude just knows what it worked on before. This is not the same as CLAUDE.md's manual approach, and it's not an MCP server — it's hook-native, which means it works wherever Claude Code's hook system works. The repo sits at ~81,000 stars as of June 2026, making it one of the fastest-growing Claude Code community projects. The project has its own docs at cmem.ai and an npm package, but the correct install path is via npx, not npm install -g.

Best for: Long-running projects where you want Claude to remember architectural decisions, debugging paths, and client context without manually curating CLAUDE.md.
Limitations: Requires Bun runtime. Memory quality depends on what Claude produces during sessions — noisy sessions produce noisy memory. Still a community project; the hooks API changed with Claude Code 2.1.0 and broke some setups.
Cost: Free, open source.

Install:

npx claude-mem install

Do not use npm install -g claude-mem — that installs only the library, not the hooks.

---

humanizer — Writing Hygiene for AI-Generated Copy

There are now several humanizer-type skills in the Claude ecosystem. The two most-referenced are blader/humanizer by Siqi Chen (~23,000 stars) and Aboudjem/humanizer-skill, which extends the original with 43 named AI writing patterns and five voice profiles.

The premise is accurate but narrower than the marketing framing suggests: these skills don't make writing "sound like a person" in any deep stylistic sense — they remove the most detectable surface patterns of AI generation. That means flagging em-dash overuse (a reliable AI tell when paired with promotional rhythm), passive voice clusters, the rule-of-three structure that Claude defaults to constantly, filler phrases ("it is worth noting that"), AI vocabulary words ("delve," "tapestry," "multifaceted"), and inflated symbolism. Aboudjem's version runs a 0–100 AI-tell score before and after rewriting, which gives you a concrete measure of improvement rather than a vague "sounds more human" claim. Three operating modes — detect, rewrite, edit — let you use it as an audit tool or a full rewrite pass. Neither version defeats trained AI detectors reliably; Turnitin and GPTZero have moved beyond surface pattern matching. Use this for editorial quality, not detection evasion.

Best for: Marketing copy, blog posts, and client-facing documentation generated with Claude where you want to reduce obvious AI tells before a human editor reviews.
Limitations: Won't beat modern AI detectors. Voice profiles are generic — you'll still need a human to inject actual brand voice. Not a substitute for editing.
Cost: Free, open source (both repos).

Install (Aboudjem/humanizer-skill, user-scoped):

mkdir -p ~/.claude/skills/humanizer && curl -sL https://raw.githubusercontent.com/Aboudjem/humanizer-skill/main/skills/humanizer/SKILL.md -o ~/.claude/skills/humanizer/SKILL.md

Install (blader/humanizer):

mkdir -p ~/.claude/skills && git clone https://github.com/blader/humanizer.git ~/.claude/skills/humanizer

---

marketingskills — A Useful Bundle, Not a Marketing Department

The claim "Claude runs your whole marketing" is overclaimed. What coreyhaines31/marketingskills actually delivers is a structured collection of 20+ marketing skills — discrete instruction sets for specific tasks — organized by CRO, copywriting, SEO, paid ads, email, and growth engineering. Corey Haines runs Swipe Files and Conversion Factory, and the repo reflects real marketing practitioner frameworks rather than generic AI marketing prompts.

The skill list is substantive: seo-audit, ai-seo, cro, signup, onboarding, copywriting, copy-edit, cold-email, ads, ad-creative, ab-testing, analytics, churn-prevention, referrals, launch, pricing, and a marketing-plan skill added in v2.0 that generates AARRR-structured plans. At ~32,700 stars, this is one of the most-starred non-Anthropic Claude skill packs. The install uses npx skillkit, which requires Node 18+. You can install the full bundle or pick individual skills. Several forks exist — syntax-syndicate/marketing-skills is a notable one — but the canonical source is coreyhaines31.

Best for: Founders and solo marketers who want Claude to produce structured marketing outputs (landing page copy, email sequences, ad creative briefs, SEO audits) with consistent frameworks, not blank-sheet prompts.
Limitations: Skills produce structured outputs, not final deliverables. You still need to review copy, verify SEO recommendations against live data, and adapt ad briefs to platform specs. Claude can't A/B test, deploy, or measure on its own.
Cost: Free, open source.

Install (all skills):

npx skillkit install coreyhaines31/marketingskills

Install (specific skills):

npx skillkit install coreyhaines31/marketingskills --skill cro copywriting

---

small-business — Anthropic's Official Back-Office Plugin

This one is different from the community builds: it's from Anthropic directly. The anthropics/knowledge-work-plugins repo contains 11 open-source plugins built for Claude Cowork (claude.com), and the small-business subfolder is also available at claude.com/plugins/small-business.

The plugin ships 31 skills organized into five operational categories: payroll and compensation, invoicing and accounts receivable, HR and onboarding, scheduling, and general operations. The skills are task-specific: /plan-payroll forecasts cash and flags payroll anomalies, /close-month reconciles transactions and drafts a P&L narrative, invoice-chasing and contract-review skills run on documents you provide. The plugin supports integrations with QuickBooks, PayPal, HubSpot, Stripe, Square, DocuSign, Gmail, Outlook, Canva, and Slack — though "integration" here means Claude can read structured exports and draft outputs in the right format, not that it connects to APIs directly without configuration. The Anthropic announcement frames this for SMBs with fewer than 50 employees who can't afford a full finance or HR stack. That framing is honest: the plugin is a structured assistant layer, not an ERP system.

Best for: Small business owners and operators who want Claude to help with invoice drafting, payroll review, month-end close narratives, and hiring packets — particularly those already on the claude.com subscription.
Limitations: This is a Claude Cowork plugin, not a Claude Code skill — the install path is different. Requires a Claude Pro or Team plan. No direct API integrations out of the box; you bring your own exports.
Cost: Plugin is free; requires Claude Pro ($20/month) or Team ($30/user/month) subscription.

Install (claude.com/plugins path):

claude plugin marketplace add anthropics/knowledge-work-plugins
claude plugin install small-business@knowledge-work-plugins

After install, run the onboarding skill by saying "set me up" — it configures the plugin to your business context.

---

Firecrawl — The Serious Web Scraping Layer

Firecrawl is not a Claude Skill — it's a standalone scraping and crawling API with a first-class MCP integration. The distinction matters: mendableai/firecrawl is a full product with cloud hosting, API keys, and pricing tiers. The claim "converts whole sites into LLM-ready content" is accurate in scope. Firecrawl takes a URL, crawls all reachable pages, renders JavaScript, and returns clean markdown or structured data. It handles authentication flows, rate limiting, and anti-bot mitigations better than most open alternatives. The main repo is around 48,000 stars; the official MCP server that connects it to Claude Desktop and Claude Code has ~5,200 stars.

For developers building research agents or data pipelines, Firecrawl sits at the right layer: you get a reliable scraping backend without building your own Playwright/Puppeteer infra, and the MCP server exposes scrape, crawl, search, map, and extract tools directly to Claude. Combine it with open_deep_research from Part 1, and you have a research loop that can pull live web content, crawl documentation sites, and extract structured data from competitor pages — all inside a Claude session. Pricing starts with a free tier of 500 credits/month (1 credit = 1 page), Hobby at $19/month for 3,000 credits, Standard at $99/month for 100,000 credits, and Growth at $399/month for 500,000 credits. Self-hosting is supported via Docker.

Best for: Research agents, competitive intelligence workflows, documentation scrapers, and any Claude pipeline that needs live web content at scale.
Limitations: Not free at meaningful volume — 500 credits/month runs out fast on a crawl. JavaScript-heavy SPAs occasionally need extra configuration. Rate limits apply at all tiers.
Cost: Free tier (500 credits/month); Hobby $19/month; Standard $99/month; Growth $399/month. Self-host option available.

Install (MCP config for Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "mcp-server-firecrawl": {
      "command": "npx",
      "args": ["-y", "firecrawl-mcp"],
      "env": {
        "FIRECRAWL_API_KEY": "YOUR_API_KEY"
      }
    }
  }
}

Get your API key at firecrawl.dev/app/api-keys. Alternatively, install via Smithery:

npx -y @smithery/cli install @mendableai/mcp-server-firecrawl --client claude

---

Repomix — Pack the Repo Slice Claude Actually Needs

If you've ever pasted an entire codebase into Claude's context window and watched it struggle with irrelevant files, yamadashy/repomix is the fix. It packs a repository — or a filtered slice of it — into a single AI-friendly file in XML, Markdown, or plain text format. XML is the default and the recommended format for Claude, which handles structured tagged context better than raw concatenation. Repomix also ships an MCP server mode, letting Claude trigger repomix operations directly during a session.

At ~26,000 GitHub stars, this is the most-starred repo-to-LLM tool in the category. The key practical feature is filtering: you can include only specific directories, exclude test files and build artifacts, and set token budget targets. For a monorepo where you only need the auth/ and billing/ modules, repomix outputs exactly those files in a structured format, with a summary header listing what's included. Pair this with open_deep_research from Part 1 when you're researching a third-party library — run repomix on the library's repo, feed the output to Claude, and get analysis that's grounded in actual source rather than training data. The output file (default: repomix-output.xml) is designed to be pasted directly into any LLM's input, not just Claude.

Best for: Code review, library analysis, architecture Q&A, and any Claude session where you need to give Claude a defined slice of a codebase without blowing the context budget on irrelevant files.
Limitations: Packing large repos with no filters can exceed context limits anyway — you still need to think about what to include. Not a RAG system; it's a one-shot file generation tool.
Cost: Free, open source. No account required for CLI use.

Install (no-install run):

npx repomix

Install globally:

npm install -g repomix

Pack only a specific directory:

npx repomix --include "src/auth/**,src/billing/**" --output auth-billing.xml

Run as MCP server (add to Claude MCP config):

{
  "mcpServers": {
    "repomix": {
      "command": "npx",
      "args": ["-y", "repomix", "--mcp"]
    }
  }
}

---

How These Stack on Top of Part 1

Part 1's five picks (Clay, taste-skill, open_deep_research, OpenBB, exo) were mostly about extending Claude's reach into data and external systems. Part 2 fills in the operational layer — context management, writing quality, and business function coverage.

The chains worth building:

Repomix + open_deep_research: Pack a third-party library with repomix, then feed the output to open_deep_research for a structured analysis of its architecture, security posture, or API surface. More grounded than asking Claude about a library from training data alone.

Firecrawl + OpenBB: Firecrawl crawls company sites, press releases, and regulatory filings into clean markdown. OpenBB pulls structured market data. Combined in a Claude session, you get a research assistant that can cross-reference unstructured web content with structured financial data — useful for due diligence, competitive analysis, or sector research.

Caveman + claude-mem: Caveman reduces what Claude writes back; claude-mem tracks what it worked on. Together they push session efficiency up at both ends — less output token spend, and future sessions don't waste tokens re-establishing context that was already established.

marketingskills + small-business: If you're a solo founder running both product and go-to-market, these two bundles cover most of what you'd otherwise break into separate workflows. marketingskills handles the outward-facing copy and growth work; small-business handles the back-office operations.

One honest note: the humanizer skills remain the most situational of the seven. They're editing tools, not voice generators. If you need consistent brand voice across Claude outputs, you'll get more mileage from a well-crafted system prompt with example text than from running a generic humanizer pass.

---

What's Coming in Part 3

The Claude Skills ecosystem is still growing faster than any curation effort can track. For Part 3, the candidates worth investigating:

Code quality and review skills — there are several skill packs focused on code review frameworks, security auditing, and test generation that have hit meaningful star counts without much coverage outside developer circles.

Diagram and architecture skills — Mermaid, PlantUML, and architecture-as-code generation skills are proliferating; the question is which implementations are actually reliable.

Voice and meeting integration — skills that connect Claude to meeting transcripts, Notion, and voice note pipelines. Several repos are iterating on this and the tooling is stabilizing.

The official Anthropic plugin library — the anthropics/knowledge-work-plugins repo ships plugins beyond small-business (sales, operations, productivity, engineering). Part 3 will go deeper on the engineering and operations plugins specifically.

If you installed any of Part 1 or Part 2's picks and ran into integration friction or found a better alternative, send it through. The most useful Part 3 inputs come from people who've actually run these in production.

---

Comparison Table

Tool	Category	Open Source?	Stars (Jun 2026)	Pricing	Claude Integration Path	Best For
Caveman	Token compression	Yes	~70k	Free	Claude Code Skill (SKILL.md)	Reducing output token costs in agentic loops
claude-mem	Session memory	Yes	~81k	Free	Claude Code hooks (5 lifecycle hooks)	Persistent cross-session context
humanizer / humanizer-skill	Writing quality	Yes	~23k (blader)	Free	Claude Code Skill (SKILL.md)	Removing surface AI tells from copy
marketingskills	Marketing	Yes	~32k	Free	Claude Code Skill (skillkit install)	Structured marketing tasks for founders
small-business	Business ops	Yes (Anthropic)	N/A (official)	Free plugin; Claude Pro required	Claude Cowork Plugin (claude.com/plugins)	Back-office ops for SMBs
Firecrawl	Web scraping	Yes	~48k	Free (500 credits); paid from $19/mo	MCP Server (npx firecrawl-mcp)	LLM-ready web content at scale
Repomix	Repo context	Yes	~26k	Free	npx / MCP server mode	Feeding filtered codebase slices to Claude