Skills are reusable, on-demand scripts and instructions that agents can call to execute specific workflows efficiently. Unlike MCPs which provide external system access, skills give agents working code for multi-step workflows like generating documents, spreadsheets, and presentations. The pattern enables 10x token reduction on subsequent runs while producing more consistent, reliable outputs.
Skills are reusable, on-demand instructions and scripts that show an agent how to perform a specific workflow efficiently. The key distinction is "on-demand" - skills are not part of the agent's system prompt but are pulled in only when needed.
A skill is typically a small folder in your repo containing everything the agent needs to execute a specific workflow. This usually includes both instructions and code, allowing the agent to reuse working code instead of reinventing it from scratch on every request.
Not a replacement for MCPs: MCP remains the easiest way to give agents access to external systems like Google Drive, Notion, Slack, Gmail, or databases. Skills can layer on top of MCPs but don't replace them.
Not a replacement for function calling: For simple, single-action tasks with known schemas and outputs (like submitting a Jira ticket or fetching a record), classic tool calling is faster, more predictable, and easier to debug.
Skills excel when there's a sequence of steps, edge cases and data transformations, iteration on output, or when the agent must generate actual artifacts (docs, spreadsheets, presentations).
Traditional tool calling breaks down here because most tools are too narrow - they expose only a few operations, boxing the agent in. With skills, instead of giving an agent 50 tools for every formatting operation, you give it one code execution tool plus a skill containing working code.
Wall 1: Token Bloat - With classic tool calling, agents need to see all tools upfront and pay the context cost for everything. With MCP servers, connecting a few can burn tens of thousands of tokens before work begins.
Wall 2: Complex Workflows - Some workflows require multiple tool calls in a specific sequence. Without saving that knowledge as a skill, the agent repeats the same mistakes on every run.
Debugging visibility: With skills, traces often show massive code chunks rather than discrete tool calls, making it harder to identify issues.
Safety best practice: Agents can use skills but cannot modify the skills directory. Lock that folder with permissions in your sandbox. Agents can generate new skills elsewhere, but they cannot override trusted, reviewed skills.
Each skill follows a specific structure with a skill.md file containing name, description (for agent discovery), and instructions for using scripts. The folder can also include executable scripts and additional markdown files for examples.
The easiest way to build skills today is with Claude Code, as Anthropic's models have been specifically trained to use skills. Anthropic provides a "skill creator" skill in their skills GitHub repo. Key advice: don't just prompt AI to create skills without context - provide your own documentation, SOPs, and process details.
"Skills are reusable on-demand scripts and instructions an agent can call to run a specific workflow efficiently. They're not just MCPs and they're not just normal tool calling either. They're the layer that makes agents compound instead of starting from scratch every single time."
"The first run used over 76,000 tokens and then on the next run after saving it as a skill, it dropped to around 8,000 tokens. Basically a 10x reduction while being a lot more reliable."
"Skills only become a powerful concept when there is code. Because without code, you're not really changing what the agent can do. You're just changing how it thinks."
"ChatGPT has already adopted skills as well without even announcing this yet, which definitely means that this is the next big thing."