This tutorial shows how to use skills to keep agent instructions modular and context-window lean. Instead of a single monolithic system prompt, skills are SKILL.md files loaded on demand when the agent decides they are relevant.
| Information | Details |
|---|---|
| Strands Features | AgentSkills plugin, Skill dataclass, @tool decorator |
| Agent Pattern | Single agent with skill-based instruction routing |
| Tools | Custom mock tools, file_read from strands-agents-tools |
| Model | Amazon Nova Lite on Amazon Bedrock |
- Skills are stored as
SKILL.mdfiles in subdirectories underskills/ AgentSkillsscans the directory and injects a compact<available_skills>summary into the system prompt- When a user message matches a skill's domain, the agent calls the built-in
skillstool to load the full instructions - The skill can reference additional files (e.g. checklists, escalation guides) via
file_read
- Python 3.10 or later
- AWS account with Amazon Bedrock access configured
- Basic familiarity with Strands Agents — see 01-first-agent if needed
15-skills/
├── README.md
├── requirements.txt
├── agent-skills.ipynb
└── skills/
├── returns-policy/
│ ├── SKILL.md
│ └── references/
│ └── returns-checklist.md
└── technical-troubleshooting/
├── SKILL.md
└── references/
└── escalation-guide.md
| File | Description |
|---|---|
| agent-skills.ipynb | Step-by-step notebook covering skill creation, plugin setup, and runtime skill injection |
SKILL.mdstructure: front-matter (name,description,allowed-tools) + instruction bodyAgentSkillsplugin: how it scans a directory, injects a skill summary, and loads full instructions on demand- Reference files: how skills use
file_readto load supporting documents at activation time - Programmatic skills: creating a
Skill(...)object at runtime without touching the filesystem - Runtime updates: using
set_available_skillsto add or replace skills mid-session
Install the required dependencies:
pip install -r requirements.txtThen open agent-skills.ipynb from the 15-skills/ directory so that relative paths to ./skills/ resolve correctly.
- 02-tools-and-mcp — building the tools that skills rely on
- 06-memory — combining skills with persistent memory
- 13-human-in-the-loop — hooks and agent lifecycle events