Course Notes · Agentic AI

Agent Skills with Anthropic

My notes from the DeepLearning.ai short course taught by Elie Schoppik. Skills are one of the most practically useful ideas I've come across in agentic AI — a clean answer to the question of how you make agents reliably do the same complex thing, over and over again.

Agentic AI Claude Skills Agents · Instructor: Elie Schoppik  ·  May 11, 2026

00 Introduction

This course covers how Agent Skills work, best practices for creating them, and how to build skills across different use-cases — from marketing automation to code review to multi-agent research pipelines.

By the end, you'll be able to:

📦 Create skills from scratch 🔗 Combine skills with MCP & sub-agents 📊 Build data analysis workflows 🧪 Review & test code with skills 🔍 Build research agents with the Agent SDK

01 What are Skills?

Skills are folders of instructions that extend your agent's capabilities with specialised knowledge or new capabilities.

A skill lives on the filesystem as a directory with a mandatory SKILL.md file and optional supporting directories for scripts, references, and assets.

📁 my-skill/
my-skill/ ├── SKILL.md # Required: metadata + instructions ├── scripts/ # Optional: executable code ├── references/ # Optional: documentation ├── assets/ # Optional: templates, resources └── ... # Any additional files

SKILL.md — the entry point

Every skill begins with a SKILL.md file containing a YAML frontmatter block with a name and description, followed by the main instructions in markdown. The instructions can reference additional files in references/ or scripts in scripts/.

📄 SKILL.md
--- name: analyzing-marketing-campaign description: Analyze weekly marketing campaign performance across channels. Use when analyzing multi-channel digital marketing data to calculate funnel metrics and compare to benchmarks, compute cost and revenue efficiency metrics, or get budget reallocation recommendations. --- # Marketing Campaign Analysis Automated analysis of multi-channel marketing campaign data ## Input Excel/CSV with columns: **Date, Campaign_Name, Channel, Segment, Impressions, Clicks, Conversions, Spend, Revenue, Orders** ## Funnel Metrics (per channel) - **CTR** = Clicks / Impressions × 100 - **CVR** = Conversions / Clicks × 100 ## References Reference to `references/budget_reallocation_rules.md`

Naming conventions

The folder name should match the name field in the frontmatter. Use lowercase letters separated by dashes (e.g. analyzing-marketing-campaign), and avoid reserved keywords like claude, anthropic, or chatgpt.

How to use a skill?

To use a skill, an agent needs two basic tools: filesystem access (to read and write files) and a bash tool (to execute code). These let the agent execute whatever a skill requires. From there, skills can be combined with MCP servers and sub-agents to build powerful agentic workflows.

For example, an agent can pull data via MCP and then rely on a skill to determine what to do with that data — how to compute metrics, apply rules, and format outputs. It can also delegate tasks to sub-agents with isolated context, which can themselves use skills for specialised knowledge.

02 When to Use a Skill?

Skills are the right tool when you have any of these situations:

Weekly / Daily
Repeated workflows you explain to the agent every time
Large Context
Need to load additional documents that fill the context window
Company Logic
LLM doesn't know how your team or company operates

In short: if you're repeatedly explaining the same workflow or bundling the same reference files every session, package it as a skill once and stop paying that tax on every conversation.

Uploading a skill to Claude

Package the skill folder into a .zip file and upload it via Capabilities → Add → upload zip file in Claude. Skills are also applicable in other coding environments like Codex, Gemini CLI, and more — they're an open standard.

03 How Skills Work

Skills are an open standard — they're transferrable across any skill-compatible agent. Build a skill once and deploy it across multiple agent products.

Progressive Disclosure

This is the core design principle behind skills. The idea is to only load data when it's actually needed — avoiding unnecessary pollution of the context window.

More data in context → more tokens consumed → faster context fill-up → greater likelihood of context degradation → incorrect responses.

Skills solve this through three loading tiers:

Always
The skill's name and description from the YAML frontmatter. These live in the agent's context window so it knows what skills are available and how to trigger them.
On Match
When a user request matches the skill's description, the full SKILL.md instruction body is loaded — the step-by-step workflow, formulas, output format, etc.
On Demand
Scripts, reference files, and assets (templates, images) are loaded separately, only if the instructions call for them. This keeps the context lean.

This architecture means you can have many skills registered at once without bloating the context window — only the skill that's actually needed gets fully loaded.

04 Why Use Agent Skills?

When building purpose-specific agents — finance, research, coding, marketing — the underlying foundation is actually a general-purpose agent: a simple scaffolding of bash and filesystem tools. What these agents lack is the underlying context and domain expertise to do the job reliably.

Old model
Research Agent
Coding Agent
Finance Agent
Marketing Agent
New model
General Agent
+
Skills
+
MCP / Sub-agents

Beyond domain knowledge, skills also provide repeatable workflows. In a non-deterministic environment where model outputs can vary, skills give the agent articulate, step-by-step instructions for a task — making agent behaviour more predictable and measurable.

Three use-case categories

Domain Expertise
  • Brand guidelines and templates
  • Legal review processes
  • Data analysis methodologies
Repeatable Workflow
  • Weekly marketing campaign review
  • Customer call prep workflow
  • Quarterly business review
New Capabilities
  • Creating presentations
  • Generating Excel / PDF reports
  • Building MCP servers

Without skills — the cost

Every session without a skill means you have to:

📝 Describe your instructions and requirements from scratch
📎 Bundle all reference files and supporting documents manually
🎲 Hope the outputs are consistent — with no guarantee they are

Skills are portable

Agent skills are now an open standard, adopted by a growing number of agent products. The same skill works in Claude Code, agents built with the Claude Agent SDK, Gemini CLI, Codex, and more.

Skills are composable

You can chain multiple skills to build complex, predictable pipelines. Here's an example marketing pipeline:

BigQuery Skill
Provide marketing-related schema & data access
Campaign Analysis Skill
Analyze funnel metrics & compute ROAS
Brand Skill
Apply brand fonts, colors, logo guidelines
PowerPoint Skill
Generate a branded slide deck

05 Skills vs Tools, MCP & Sub-agents

Skills vs MCP

🔌 MCP

Connects your agent to external systems and data — Google Drive, databases, APIs, and more.

MCP is the pipe that brings data in.

🧠 Skills

Teaches your agent what to do with that data — how to analyse it, which rules to apply, how to format the output.

Skills are the expertise that acts on the data.

Skills vs Tools

A good analogy: tools are a hammer, saw, and nails — the raw primitive operations. A skill is the knowledge of how to build a bookshelf — the methodology that orchestrates those tools toward a specific outcome.

Tool definitions always live in the model's context window. Skills are progressively loaded only when needed — this distinction is key. Skills can also include tools on demand, triggered selectively through progressive disclosure rather than always occupying context.

Skills vs Sub-agents

A main agent can spawn sub-agents that work independently and report back. Sub-agents have isolated context with fine-grained permissions and can run tasks in parallel. You can specify which skills each sub-agent has access to, letting the main agent act as orchestrator while sub-agents handle specialised work.

Main Agent (Orchestrator)
Interview Analyzer
skills + tools
Survey Analyzer
skills + tools
Code Reviewer
skills + tools

Summary comparison

Feature Skills Prompts Sub-agents MCP
What it provides Procedural knowledge Moment-to-moment instructions Task delegation Tool connectivity
Persistence Across conversations Single conversation Across sessions Continuous connection
Contains Instructions + code + assets Natural language Full agent logic Tool definitions
When it loads Dynamically, as needed Each turn When invoked Always available
Best for Specialised expertise Quick requests Specialised tasks Data access

06 Pre-built Skills

Anthropic ships a set of pre-built document skills you can use out of the box. These encode hard-won best practices for creating each document type — available libraries, rendering quirks, output conventions.

.pptx
Presentations & slide decks
.docx
Word documents & reports
.pdf
PDF creation & form filling
.xlsx
Spreadsheets & models

There's also a skill creator skill — a meta-skill that guides the agent through the creation process and best practices for writing new skills. It's one of the most useful tools in the kit for bootstrapping domain-specific skills quickly.

Hands-on exercises in the course

1. Use the skill creator to build a BigQuery skill
2. Create a company brand guidelines skill (with logos)
3. Combine brand skill + PowerPoint skill for branded decks
4. Build a research agent with the Claude Agent SDK

07 Key Takeaways

Agent skills are genuinely one of the most practical primitives I've seen in agentic AI. The core insight is simple: don't explain, package. If you're repeating yourself to an agent — whether that's workflow instructions, domain rules, or reference documents — wrap it in a skill and let the agent discover it automatically.

The progressive disclosure model is elegant. By keeping only the skill's name and description in context at all times, you can register many skills without bloating the context window. The full instructions only load when they're actually needed — a smart tradeoff between discoverability and efficiency.

And the open-standard angle matters: a skill written today for Claude Code works tomorrow in any other skill-compatible agent environment. The investment compounds.

The analogy that stuck with me: tools are the hammer, saw, and nails. A skill is knowing how to build a bookshelf. The general-purpose agent already has the tools — it just needs the expertise.