Skills.
Give your agents knowledge — without the bloat
Skills are on-demand instructions agents load only when relevant, so your context window stays clean and focused.
How agents use skills at runtime
The agent sees all available skill names and short descriptions appended to its system prompt.
The agent evaluates which skills, if any, are relevant to the current task input.
The agent fetches additional instructions, only when relevant.
The agent follows the loaded instructions and optionally fetches linked reference files for deeper detail.
Skills in Kiln
Context-efficient by design
Each skill injects only its name and description — about 100 tokens — into the system prompt. Full instructions stay out of context until the agent explicitly requests them, keeping your context window available for actual work.
Nested files for deeper knowledge
Skills can nest additional knowledge and instructions. The agent navigates this tree on demand — loading only what's needed for this task.
Skill versioning
Each skill is versioned, and tracked in your agent's run configuration. This keeps the eval history clean and experiments traceable as you iterate on skills.
Everything you need for Skills
Full instructions fetched only when the agent calls skill().
Only name and description added to the system prompt per skill.
Link to nested docs for multi-level knowledge hierarchies.
Create once, use across any task in the project.
New versions via clone; originals stay intact for eval validity.
Skill calls appear in the All Messages trace with full detail.
Choose which skills are available per run under Advanced > Skills.
Skills stored on your filesystem. MIT-licensed Python library, source-available app.
Skills vs. tools vs. RAG
Three ways to extend your agent—each built for a different job. Kiln can find the best one for your agent.
| Capability | Skills | Tools (MCP) | RAG (Search Tools) |
|---|---|---|---|
| Primary use | Knowledge & instructions | External actions | Large-scale document search |
| Context overhead | ~100 tokens per skill | Tool definitions | 1 tool definition for all docs |
| Progressive disclosure | — | ||
| Max documents | ~50 | — | Unlimited |
| Agent control over retrieval | High | — | Low |
| Setup complexity | Low | Medium | High |
| Best for | Guidelines, procedures, standards | APIs, databases, file systems | Large knowledge bases |
Frequently asked
How does the agent decide which skill to use?
When skills are enabled, each skill's name and description are appended to the agent's system prompt. The agent decides — based on task input and reasoning — which skills to load. If a skill isn't being used when it should, improve its description or mention the skill explicitly in your task prompt.
How are skills different from just putting instructions in the system prompt?
System prompt instructions are always present, consuming tokens whether relevant or not. Skills add only a short description (~100 tokens each); full instructions are fetched on demand. Your context window stays available for actual work and the model isn't overwhelmed with irrelevant guidelines.
What are reference files?
Reference files are additional documents in a references/ subdirectory inside the skill folder. They let you build nested knowledge hierarchies — main instructions can link to reference files, which can link to further references. The agent navigates the tree on demand.
Stop stuffing your system prompt. Use skills.
Give your agents focused, on-demand knowledge with skills—no infrastructure, no bloated context, no guessing.