Claude Series – Post 6: Skills - Turn Claude Into YOUR Specialist! 🧬 Link to heading

Remember CLAUDE.md from Post 5? That was about onboarding Claude into your project — telling it about your files, packages, and standards.

Skills take it further. Instead of just informing Claude about your project, you teach it how to perform specific tasks exactly the way you want — and it remembers that expertise forever.

🎯 What Are Skills? Link to heading

Skills are reusable instruction packages that extend Claude’s capabilities. Each Skill is a folder containing a SKILL.md file with YAML frontmatter (name + description) and detailed Markdown instructions. You can optionally include scripts, templates, and reference files.

The clever part is progressive disclosure. Claude scans only the metadata of all available Skills (~100 tokens each), then loads the full content of only the relevant ones when your task matches. Dozens of Skills, zero context bloat.

How Claude Uses Them Link to heading

Automatically: Claude reads Skill descriptions and loads matching ones when your task fits. Ask “filter my Seurat object for low-quality cells” and Claude auto-loads your QC Skill without being told.

Manually: In Claude Code, every Skill becomes a slash command. Type /scrnaseq-qc to force-load it, optionally with arguments:

/scrnaseq-qc Filter the PBMC dataset with strict thresholds

📂 Where Do Skills Live? Link to heading

  • Claude chat: Upload a zipped Skill folder in Settings → Capabilities → Skills
  • Claude Code (project): Drop folders into .claude/skills/ in your project — version-controlled and shared via git
  • Claude Code (personal): Place in ~/.claude/skills/ for Skills available across all projects
  • Marketplace: Install community Skills with /plugin marketplace add

Priority when names collide: Enterprise > Personal > Project.

💡 Built-In and Partner Skills Link to heading

Claude ships with professional Skills for docx, xlsx, pptx, and pdf creation. Say “take my DESeq2 results and create a Word document with summary table and volcano plot” and Claude produces a publication-ready file. The Skills Directory also features partner Skills from Notion, Figma, and Atlassian that integrate with MCP connectors (more on MCPs next post!).

🚀 Superpowers: The Community Game-Changer Link to heading

Superpowers by Jesse Vincent is a battle-tested library of 20+ Skills for Claude Code. Install it in two commands:

/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

What you get:

  • /superpowers:brainstorm — Interactive design refinement before writing code
  • /superpowers:write-plan — Structured implementation plans
  • /superpowers:execute-plan — Executes plans in verified batches

Plus auto-activating Skills that fire when Claude detects the right context: test-driven-development when implementing features, systematic-debugging when tracing bugs (root cause, not symptoms), and verification-before-completion before Claude claims work is done.

The bootstrap is fewer than 2,000 tokens — individual Skills load on demand via a search script, so you never pay for what you don’t use. The broader awesome-claude-skills repository curates even more community Skills across scientific computing, security, writing, and more.

🏗️ Creating Your Own Skills Link to heading

Step 1: Create the Directory Link to heading 

# Project-level
mkdir -p my-project/.claude/skills/scrnaseq-qc

# Or personal (all projects)
mkdir -p ~/.claude/skills/scrnaseq-qc

Step 2: Write SKILL.md Link to heading

The description is critical — it’s what Claude reads to decide relevance:

---
name: scrnaseq-qc
description: "Standard single-cell RNA-seq quality control workflow. 
Use when performing QC filtering on 10X Chromium scRNA-seq data, 
setting QC thresholds, or filtering cells by nFeature, nCount, 
and percent.mt metrics."
---

# Single-Cell RNA-seq QC Protocol

## Standard Thresholds
- nFeature_RNA: 200 - 5000
- nCount_RNA: minimum 500
- percent.mt: < 15% (tissue-dependent)

## Workflow Order
1. Run SoupX on raw matrix FIRST
2. Create Seurat object from corrected counts
3. Calculate percent.mt with PercentageFeatureSet(pattern = "^MT-")
4. Visualize QC metrics BEFORE any filtering
5. Apply thresholds based on distribution
6. Run DoubletFinder AFTER initial filtering
7. Document all thresholds and cells removed

## Critical Rules
- NEVER skip ambient RNA removal for droplet-based data
- ALWAYS set seed (42) for reproducibility
- ALWAYS generate QC plots BEFORE and AFTER filtering

Step 3: Add Supporting Files (Optional) Link to heading 

scrnaseq-qc/
├── SKILL.md              # Main instructions (required)
├── templates/
│   └── qc_report.Rmd     # Template Claude fills in
├── scripts/
│   └── validate_qc.R     # Executable validation
└── resources/
    └── marker_genes.csv   # Reference data

Step 4: Share It Link to heading 

git add .claude/skills/scrnaseq-qc/
git commit -m "Add scRNA-seq QC skill"
git push

Everyone who pulls the repo gets the same protocol. For Claude chat, zip the folder and share the .zip file — colleagues upload through Settings.

Pro tip: Claude ships with a built-in skill-creator Skill. Just ask “help me create a new Skill for [your workflow]” and it walks you through the process interactively.

🤝 Skills vs CLAUDE.md Link to heading

CLAUDE.md = project context (the what and where): project description, directory structure, package versions, naming conventions.

Skills = task expertise (the how): how QC should be performed, how reports should be formatted, how figures should be generated.

They’re complementary. CLAUDE.md gives Claude the map. Skills give Claude the expertise to navigate it. Use both.

⚡ Tips for Effective Skills Link to heading

Write specific descriptions. “Data analysis helper” won’t trigger reliably. “Single-cell RNA-seq QC for 10X data with SoupX and DoubletFinder” will.

Keep Skills focused. One Skill per workflow. Focused Skills compose better than one monolithic Skill.

Include examples. A sample_output.md showing expected output is worth a thousand words of instruction.

Version control everything. When your lab updates a protocol, update the Skill. Everyone pulls the latest automatically.

🔥 The Bottom Line Link to heading

Skills transform Claude from a general assistant into YOUR specialist — reusable, shareable, composable, and automatically invoked. Set them up once, benefit forever. They work across Claude chat, Claude Code, and the API. Write once, use everywhere.

Next up: MCPs (Model Context Protocol) — connecting Claude to external tools, databases, and services! If Skills teach Claude how to do tasks, MCPs give Claude access to the tools and data it needs. The combination is where things get really powerful. 🔌

Full Claude series at badran-elshenawy.netlify.app/posts