Skills & Plugins — specializing Claude
A Skill = an instruction file (SKILL.md) that teaches Claude a specific competency. Like a specialization degree.
Until now, creating a Skill meant writing the method: a file, steps, a format. Anthropic flipped the burden of proof. You hit record, you do the task the way you always do while talking through it, you click Done — and Claude hands you a ready-to-use Skill built from what it saw and heard you do.
The detail that changes everything: narration. Claude sees your clicks, but it's your voice that supplies the why — why you skip a step, how you pick between two options, what you check before hitting send. A silent demo produces a macro; a narrated demo produces a method. That is exactly the gap between a robot that repeats and a colleague who understood.
Where to find it: the + button in the composer → Record a skill, or Customize → Skills → Add → Record your screen.
Why Skills change everything
🔁 Reproducibility
No need to re-explain. Claude knows your methodology.
🏢 Scalability
One expert's Skill = a tool accessible to the whole organization.
Skills by business department
Legal
Contracts, NDAs, GDPR, AI Act, M&A terms.
Marketing
SEO, editorial calendar, brand voice, B2B LinkedIn.
HR
Job descriptions, resume screening, interview frameworks, workforce planning.
Finance (CFO)
Financial model audits, LBO/DCF, variance, reconciliation.
How to acquire Skills
Cowork Marketplace: Claude Desktop → Cowork → Marketplace → Install.
Community: /plugin marketplace add jeremylongshore/claude-code-plugins
Directories: claudemarketplaces.com (2,300+), skillsmp.com (700K+), GitHub.
Build your own: a SKILL.md file < 2,000 tokens in ~/.claude/skills/.
Record it: you do the task on screen while narrating it, and Claude writes the Skill — see Record a Skill above.
The simplest trick: paste a GitHub URL to Claude
Beyond the skills natively offered by Anthropic, thousands of additional skills and plugins are published on GitHub. You don't need to know what GitHub is to benefit: copy the page's address, paste it into Claude, and ask for its take. Claude does the analysis work — and the installation if you give the green light.
Copy the URL — you spot an interesting skill or plugin (a directory, a LinkedIn article, a recommendation): just copy the GitHub link, for example github.com/anthropics/skills.
Paste it into Claude and ask — "Here's a skill I found on GitHub: [URL]. Explain what it does, how it would concretely help me in my job, and whether it carries any risk." Claude reads the page, the SKILL.md file, and gives you a clear plain-English summary.
Decide — and let Claude load it — if the analysis convinces you: "Great, install it." Claude fetches the skill and sets it up. Otherwise, you move on. You keep the decision, it does the legwork.
Pro tips & common mistakes
✅ Pro tips
→ Create your own .md files with your rules (editorial guidelines, brand, checklists).
→ One Project per recurring task. Don't build a mega catch-all Project.
→ Upload your 5 best reference examples — not 50 files dumped in bulk.
→ Type / in any chat to see all the available commands.
→ Choose 2-3 plugins max. Each one adds context = less room for your request.
❌ Common mistakes
→ Pasting your bio into every new chat = use Projects.
→ Skipping the model + settings configuration = Claude sounds generic.
→ Dumping 200 files and hoping Claude finds the right one = pick the best ones yourself.
→ Thinking Artifacts are just demos = ask for what you'd build in a spreadsheet or Canva.
→ Confusing Skills and Projects = Projects hold your files, Skills teach Claude how to perform a task.
The "context pack" — Cowork version (no technical skills)
The most powerful idea from advanced users: never re-explain who you are or how you work. Developers do this with .md files; in Cowork, you get the same result without a line of code, via three locations. Think of it as the welcome binder you'd hand a new hire.
Preferences
Who you are and how you want the answers (tone, language, level of detail, structure). Applies to all your chats. Set it once.
Project instructions
The context specific to an engagement: goal, audience, business rules, constraints, internal vocabulary. Applies to all chats in this project.
Reference files
Your 5 best examples (a memo template, brand guidelines, a sample report) placed in the Project. Claude uses them as a benchmark — not 50 files dumped in bulk.
For technical teams, the equivalent for Claude Code is called CLAUDE.md — detailed in Module 06.
Create a "Executive committee communication" Project and paste the suggested instruction block. Open a chat in this project and ask for a memo.
Expected result : The memo follows your rules without you repeating them.