What Are Skills?
Agent skills are structured prompts and knowledge that agents can discover and load on the fly. Instead of baking everything into the agent upfront, skills let it extend its own capabilities based on the task at hand.
Why It Works: Progressive Disclosure
As models get better, teams keep adding more information to their system prompts. This has a cost: both literally, and in terms of quality, as the model's ability to follow complex instructions drops off as prompts get longer.
Skills use a concept called progressive disclosure: the agent knows what information is available and loads it when needed, instead of stuffing it all into a system prompt upfront.
For example, a customer service agent might handle a dozen types of requests, but each call typically focuses on one or two. Without skills, you'd load 10,000 tokens of refund policies only to find out the caller just wanted business hours. With skills, the agent loads only the information it needs, when it needs it.
Nesting Skills
Progressive disclosure can itself be nested. Skills can load reference files (think sub-skills), forming a tree of knowledge. The agent navigates the tree to find exactly what it needs while keeping the context window focused. There's no limit to the depth of the tree, allowing you to scale progressive disclosure in log(n) time!
📂 Customer Service Skill ├── 📁 refunds/ │ ├── new-customer-no-questions.md │ ├── loyalty-winback-offers.md │ ├── chargeback-dispute.md │ └── ... ├── 📁 upgrades/ │ ├── enterprise-custom-pricing.md │ ├── trial-extension.md │ └── ... ├── 📁 suspend-account ├── 📁 billing-and-invoices ├── 📁 hours-and-contact └── ...
Compared to RAG
Skills and RAG solve similar problems differently. Which is best depends on your use case, but we're finding skills are the better fit for most of the agents we build and work with.
- Pro: Much easier to build and deploy
- Pro: More control over the content of each file/chunk
- Pro: The agent sees a summary of available knowledge and can fetch exactly what it needs, instead of guessing at keywords
- Neutral: Requires manually designing your skill tree, which gives you control but takes more effort than dumping docs into a vector store
- Con: More limited in scale. RAG can handle essentially unlimited documents, while each skill adds to your context. Nesting reference files helps, but there's still a practical ceiling.
Get Started with Agent Skills
Ready to give your agents the knowledge they need, exactly when they need it? Download Kiln to start building with agent skills today. Click the "Skills" tab in the app to add your first skill, and check out our docs for more advanced features like nested reference files.
