Loading…
Guide
Quick answer
To build a Claude skill, scope it to one clear job, write a precise trigger description so Claude knows when to use it, add readable instructions and any optional scripts, test it against real tasks, and package it (SKILL.md plus assets). Narrow, well-documented skills work best.
The best agent skills are narrow and reliable, not sprawling. Here's how to build one that triggers correctly and is safe to share.
Pick a single, well-defined capability — code review, a doc generator, one integration. Narrow skills trigger more reliably than catch-alls.
Describe exactly when the skill applies. This is what the agent reads to decide whether to load it, so be specific about the tasks it matches.
Write step-by-step instructions the agent follows, plus any optional scripts or assets. Keep them readable and auditable.
Only access the data and tools the skill truly needs, and document what it touches so users can trust it.
Run the skill on real prompts to confirm it triggers at the right time and produces correct results before sharing it.
Bundle the SKILL.md and assets, state compatibility and license, and list it on a marketplace so others can install it.
A single, clear responsibility; a precise trigger description; readable instructions and auditable scripts; minimal permissions; and documentation written for agents and humans. Narrow, reliable skills beat sprawling ones.
Production-ready kits, components, agent skills and more — with live previews and clear licensing.
Explore the marketplace