PaddySpeaks · AI Infrastructure Series · Part I of II

The Skill Set

A field guide to Claude's Agent Skills — what they are, how they work, why they matter, and how the overzealous among us keep tripping over them.

March 2026 · 18 min read · paddyspeaks.com

Every time you ask Claude to make you a PowerPoint deck, extract tables from a PDF, or build a landing page that doesn't look like every other AI-generated artifact on the internet — you're triggering a Skill. You just didn't know it had a name.

Agent Skills are Claude's quiet superpower. They're folders of markdown instructions, bundled scripts, and reference materials that Claude loads dynamically — only when it recognizes that a task matches what the Skill knows how to do. Think of them as onboarding documents for a specialist. You wouldn't hand a new data engineer your entire company wiki on day one. You'd give them the runbook for the system they're about to operate. Skills work exactly the same way.

This article is the first of a two-part series. Here, we map the full terrain of Skills — what ships with Claude, what the community has built, how to create your own, and where enthusiastic newcomers tend to stumble. In Part II, we'll cover Connectors: how Claude talks to the outside world through Google Drive, Slack, Salesforce, and dozens of other integrations.

Let's walk the catalog.

I
Section 01

What Is a Skill, Really?

At its core, a Skill is just a folder. Inside that folder is a file called SKILL.md — a markdown document with YAML frontmatter at the top (the name and description) and instructions in the body. That's the minimum. A Skill can also include scripts, reference documents, templates, fonts, and asset files.

The magic isn't the format. The magic is progressive disclosure — a three-level loading system that keeps Claude fast and focused. Click each level below to explore:

Interactive · Click Each Level
1
Metadata — Name + Description
Always loaded

The YAML frontmatter from every installed Skill's SKILL.md. Only ~50-100 tokens per Skill. Claude reads all of these at startup — this is how it knows what's available.

Think of it as the menu at a restaurant. You scan everything, but you don't eat it all.
2
SKILL.md Body — Full Instructions
On trigger

The complete instruction set. Loaded into context only when Claude decides this Skill matches the current task. Best kept under 500 lines — if you need more, point to reference files.

The recipe. You read it after you've ordered — not before you walk into the restaurant.
3
Bundled Resources — Scripts, Assets, References
On demand

Python scripts, template files, fonts, reference docs. Loaded only when the instructions call for them. Scripts can execute without being loaded into the context window — they run in the VM.

The ingredients. The kitchen fetches them when the chef is ready to cook — not when you place the order.

This is why Skills don't bloat your conversations. Claude carries the menu of what's available (Level 1), reads the recipe when you order something (Level 2), and grabs the ingredients only when it's time to cook (Level 3).

How a Skill Triggers — The Flow

What happens between your prompt and Claude's response? Here's the sequence. Hit play to watch it unfold:

Interactive · Skill Trigger Sequence
💬
Your Prompt
"Create a 10-slide investor pitch deck…"
🔍
Scan Metadata
Claude scans every Skill's name + description
🎯
Match Found
pptx Skill description matches "deck" + "slides"
📖
Load SKILL.md
Full instructions loaded into context window
Execute
Scripts run, assets load, .pptx file created

Anatomy of a Skill Folder

my-skill/
├── SKILL.md ← required. instructions + frontmatter
└── Bundled Resources ← optional
    ├── scripts/ ← executable code for deterministic tasks
    ├── references/ ← docs loaded into context as needed
    └── assets/ ← templates, icons, fonts, images

Watch Out — The Description Is the Trigger

The description field in your YAML frontmatter isn't just documentation. It's the primary mechanism Claude uses to decide whether to invoke your Skill. A vague description means Claude ignores your Skill. An overly specific one means it only fires for exact phrasing. Anthropic's own guidance: make descriptions slightly "pushy" — include the scenarios and keywords where the Skill should activate, not just what it does.

II
Section 02

The Document Skills — Claude's Bread and Butter

These are the Skills most people encounter without realizing it. Every time Claude creates a Word document, builds a slide deck, generates a spreadsheet, or fills a PDF form, it's invoking one of four core document Skills. They ship pre-installed on every paid Claude account.

📄
docx
Document Skill

Creates, reads, edits, and manipulates Word documents. Handles tables of contents, headings, page numbers, letterheads, tracked changes, find-and-replace, and image insertion. Backed by a full Python scripting pipeline.

Example Prompt

"Write me a professional memo to the engineering team about our Q2 data migration timeline. Include a table of milestones. Save it as a Word doc."

📊
pptx
Document Skill

Creates, reads, parses, edits, and combines PowerPoint presentations. Uses the PptxGenJS library under the hood. Handles speaker notes, layouts, slide masters, and comments. Triggers on any mention of "deck," "slides," "presentation," or a .pptx filename.

Example Prompt

"Create a 10-slide investor pitch deck for a seed-stage AI compliance startup. Include a market size slide and a competitive landscape slide."

📈
xlsx
Document Skill

Opens, reads, edits, and creates Excel spreadsheets. Handles formulas, formatting, charting, and cleaning messy data. Also works with .csv and .tsv files. The deliverable must be a spreadsheet file — if you just want a table in chat, this Skill doesn't trigger.

Example Prompt

"My boss sent me this Q4 sales file — it's a mess. Add a profit margin column (revenue is in C, costs in D) and create a chart showing monthly trends."

📋
pdf + pdf-reading
Document Skill

Two complementary Skills. The pdf Skill creates, merges, splits, watermarks, encrypts, fills forms, and extracts images from PDFs. The pdf-reading Skill handles reading strategy — text extraction, page rasterization, embedded content, and OCR on scanned documents.

Example Prompt

"Read this uploaded lease agreement PDF and extract all the key financial terms — monthly rent, deposit, escalation clauses — into a summary table."

The Overzealous Trap — "Make Everything a File"

New users frequently ask Claude to create a .docx for things that belong in chat. A quick summary? A three-paragraph email? A simple list? These don't need to be Word documents. Creating a docx takes significantly more time and tokens than responding inline. Save document Skills for deliverables you'll actually download, share, or edit outside Claude.

III
Section 03

The Utility Skills — Invisible but Essential

📂
file-reading
Utility Skill

A routing Skill. When you upload a file, this Skill figures out what kind of file it is and tells Claude which tool to use — not all files are equal. A .pdf needs different handling than a .csv, and a .xlsx isn't the same as a .json. This is the traffic cop that prevents Claude from running cat on a binary file and returning gibberish.

Example Prompt

"I uploaded a file. What's in it?" (The file-reading Skill activates behind the scenes to determine the right extraction strategy.)

🧠
product-self-knowledge
Utility Skill

An often-overlooked Skill that provides Claude with accurate, up-to-date facts about Anthropic's own products — Claude Code installation, API function calling, SDK usage, rate limits, plan features, pricing. Without it, Claude might hallucinate product details from its training data. Triggers on questions about Claude Code, Claude API, or Claude.ai features.

Example Prompt

"How do I set up MCP server integration with Claude Code? What's the Node.js version requirement?"

🎨
frontend-design
Design Skill

The anti-slop Skill. Without it, Claude defaults to what Anthropic calls "distributional convergence" — Inter font, purple gradients, centered layouts. This Skill forces bold aesthetic direction: distinctive typography, unexpected spatial composition, and a deliberate design philosophy before any code is written. It's the reason some Claude artifacts look designed rather than generated.

Example Prompt

"Build me a landing page for a meditation app. I want it to feel like a Japanese rock garden — minimal, textural, contemplative."

Watch Out — Skills ≠ Personality

The frontend-design Skill changes how Claude builds interfaces, not what it builds. You still need to provide direction. Saying "make it look good" is like telling an architect "make it nice." The Skill gives Claude a design philosophy framework — you still need to give it a brief.

IV
Section 04

The Example Skills — Where Things Get Interesting

Beyond the core document and utility Skills, Anthropic ships a collection of example Skills that demonstrate what's possible. Use the explorer below to browse by category. Click any card to see an example prompt.

V
Section 05

Creating Your Own Skill — The Practical Path

You don't need to be a developer. You need to be clear about what you want Claude to do repeatedly, and write it down in markdown. Here's the minimum viable Skill:

Minimum SKILL.md --- name: my-first-skill description: Formats meeting notes into a structured action-item summary. Use when the user mentions meeting notes, action items, or standup summaries. --- # Meeting Notes Formatter When given raw meeting notes, produce: 1. A three-sentence summary 2. A table of action items with owner and deadline 3. A list of open questions Use Markdown formatting. Keep the tone professional but concise.

The Five-Step Creation Loop

Click each step to expand. They build on each other — this is how Anthropic's own skill-creator Skill works under the hood.

Step 01
Capture Intent

What should this Skill enable? When should it trigger? What's the expected output format? Be specific about edge cases. The skill-creator Skill will interview you about all of this — or you can answer these four questions yourself: What does it do? When does it trigger? What's the output? Should we set up test cases?

Tip: Don't skip edge cases. "What if the input is messy?" is always a good question.
Step 02
Write the SKILL.md

Start with a slightly "pushy" description — include keywords and scenarios, not just capabilities. Keep the body under 500 lines. Use the imperative form. Explain why, not just what. If Claude understands your reasoning, it generalizes better than when you issue rigid commands.

Tip: ALWAYS and NEVER in all caps is a yellow flag. Reframe as reasoning instead.
Step 03
Test with Real Prompts

Write 2-3 realistic test prompts — the kind of thing someone would actually type. Include casual phrasing, typos, abbreviations. "ok so my boss sent me this xlsx file called Q4 sales final FINAL v2.xlsx" is a better test case than "Format this data."

Tip: If your test prompts are perfect English, they're not realistic enough.
Step 04
Evaluate and Iterate

Did the Skill trigger when it should have? Was the output format right? Were there edge cases that broke things? Adjust the description, refine the instructions, run again. The skill-creator includes an eval viewer that shows outputs side-by-side with feedback textboxes.

Tip: Empty feedback means it looked fine. Focus improvement energy on specific complaints.
Step 05
Optimize the Description

Create 20 eval queries — half should-trigger, half shouldn't. The negative cases are the hard part: they should be near-misses that share keywords but need a different Skill. The optimizer splits into train/test sets, runs 5 iterations, and picks the best description by test score to avoid overfitting.

Tip: "Write a fibonacci function" is a useless negative case for a PDF Skill. Make negatives tricky.

Installing Your Skill

PlatformHow to Install
Claude.aiSettings → Capabilities → Skills → Upload your Skill folder (zipped if needed)
Claude CodePlace in ~/.claude/skills/ or install via npx skills add
Claude APIUpload via the /v1/skills endpoint or include in Messages API requests

The Overzealous Trap — "I'll Skill Everything"

Not every workflow needs to be a Skill. Skills are for repeatable patterns you want Claude to execute consistently across conversations. A one-off task? Just prompt it. A process you've explained three times this week? That's a Skill.

Also: Skills don't currently sync across surfaces. A Skill uploaded to Claude.ai isn't automatically available in Claude Code or the API. You'll need to install it separately on each platform.

VI
Section 06

Prompting for Skills — The Cheat Sheet

Skills trigger based on task context, not magic words. But there are patterns that reliably activate the right Skill — and patterns that confuse the routing.

Prompts That Trigger Well

Document Creation "Create a Word document with our Q3 marketing results. Include a table of contents and executive summary." "Build a 12-slide deck on our product roadmap. Use speaker notes for the presenter." "Take this messy CSV and turn it into a formatted Excel spreadsheet with charts."
File Processing "Read this PDF and extract all the financial terms into a structured table." "Merge these three PDFs into one document, add page numbers, and watermark each page with 'DRAFT'." "This uploaded Excel file has duplicate rows and missing headers. Clean it up."
Design & Creative "Build me a landing page for a wellness brand. I want it brutalist — raw, typographic, no stock photo energy." "Create a piece of generative art inspired by the mathematical patterns in sunflower spirals." "Design a poster for a jazz festival. Think Blue Note album covers — bold, graphic, minimal."

Test Yourself — Will It Trigger?

Interactive Quiz
Does this prompt trigger a Skill?

1 / 6

Prompts That Don't Trigger (and Why)

PromptWhy It FailsFix
"Give me a table of Q3 results"No file keyword — Claude returns a markdown table in chatAdd "as an Excel file" or "as a spreadsheet"
"Summarize this document"Too simple — Claude handles this without a SkillOnly needs a Skill if the output should be a file
"Make it look nice"No design direction — frontend-design Skill can't commit to an aestheticGive a tone, reference, or metaphor: "like a Japanese garden"
"Read file X"Too simple and one-step — Claude handles basic file reads directlyWorks fine for basic reads; Skills activate for complex multi-step tasks
VII
Section 07

Seven Watchouts for the Overzealous

Enthusiasm is good. Informed enthusiasm is better. Here's where most people trip.

1. Security is your responsibility. A Skill can include executable scripts. A malicious Skill can direct Claude to exfiltrate data or make unauthorized API calls. Only use Skills from trusted sources — ones you wrote yourself or obtained from Anthropic. If you're auditing a third-party Skill, read every file in the folder, not just the SKILL.md.

2. Skills are not prompts. Writing "ALWAYS DO X" in all caps doesn't make a good Skill. Anthropic's own Skill-creation guidance says: explain the why, not just the what. Claude has good theory of mind. When you explain reasoning, it generalizes better than when you issue commands.

3. Description overfitting is real. If your Skill triggers only for the exact phrases you tested, you've overfit the description. The skill-creator's optimizer combats this by splitting eval queries into train/test sets — but you need to write genuinely tricky negative cases, not obviously irrelevant ones.

4. Context window is finite. Every Skill loaded into context consumes tokens. If you have 30 Skills installed and they all trigger at once, you've eaten into the space available for your actual task. Keep SKILL.md files lean (under 500 lines). Move reference material to separate files that load on demand.

5. Skills don't compose automatically. While Anthropic describes Skills as "composable" — Claude can use multiple Skills in one task — this works best when Skills have distinct domains. Two Skills that both want to format the same output will conflict. Design for orthogonality.

6. The Agent Skills standard is cross-platform, but… The same SKILL.md format works across Claude Code, Cursor, Gemini CLI, Codex CLI, and other compatible agents. However, Skills designed to use Claude-specific tools (like the VM filesystem or Artifacts) won't work everywhere. Use the compatibility field in your frontmatter to note platform requirements.

7. Undertriggering is more common than overtriggering. Anthropic's own documentation acknowledges that Claude has a tendency to not use Skills even when they'd help. The fix is in the description — make it clear, keyword-rich, and slightly pushy about when to activate.

VIII
Section 08

Glossary

Agent Skill
A folder containing a SKILL.md file (and optionally scripts, references, and assets) that extends Claude's capabilities for a specific task domain.
SKILL.md
The required markdown file in every Skill folder. Contains YAML frontmatter (name, description) and the body instructions Claude follows.
Frontmatter
The YAML metadata block at the top of SKILL.md, delimited by ---. The description field here is the primary trigger mechanism.
Progressive Disclosure
The three-level loading system: metadata always in context, SKILL.md body loaded on trigger, bundled resources loaded on demand.
Triggering
The process by which Claude decides a Skill is relevant to the current task, based on the description field. Simple one-step tasks may not trigger Skills even if the description matches.
MCP
Model Context Protocol. An open standard for connecting LLMs to external services. MCP servers are the bridges; Skills like mcp-builder help create them.
Connector
An MCP integration that lets Claude interact with external services (Google Drive, Slack, Salesforce). The subject of Part II of this series.
Subagent
An independent Claude instance spawned to run a specific task (like testing a Skill). Available in Claude Code and Cowork but not in Claude.ai's web interface.
Distributional Convergence
Anthropic's term for why AI outputs tend to look the same — models reproduce the statistical center of their training data. The frontend-design Skill explicitly combats this.
.skill File
A packaged Skill ready for distribution. Created by the skill-creator's packaging script. Can be uploaded to Claude.ai or shared with team members.
IX
Section 09

Further Reading

Part 2 of 2

The Skills People Actually Use

Which Skills get invoked most in production? How do teams compose Skills for complex workflows? In Part 2, we look at real-world usage patterns, the Skills that ship with Claude Code, and the community-built Skills that are quietly changing how people work.