Skills Configuration
Skills provide specialized instructions and techniques that agents can follow. AgentPool supports automatic injection of skills into agent system prompts using structured XML formatting.
Overview¶
Skills injection allows you to:
- Automatically include relevant skill instructions in agent prompts
- Configure global defaults for all agents
- Override per-agent using skills tool configuration
- Limit the number of skills included to manage token count
Configuration Structure¶
Global Skills Configuration¶
skills:
# Skill discovery paths
paths:
- ~/.config/agentpool/skills
- ./skills
# Include default AgentPool skills (default: true)
include_default: true
# Instruction injection configuration
instruction:
# Injection mode: off, metadata, or full
mode: metadata
# Maximum number of skills to inject (default: 20)
max_skills: 20
Injection Modes¶
| Mode | Description | Use Case |
|---|---|---|
off |
No skill injection (default) | Disable automatic injection |
metadata |
Skill names and descriptions only | Quick reference without full content |
full |
Complete skill instructions | When skills contain critical instructions |
Agent-Specific Overrides¶
Override global settings for specific agents using the skills toolset:
agents:
expert_coder:
model: openai:gpt-4o
system_prompt: "You are an expert developer"
tools:
- type: skills
# Override global injection mode for this agent
injection_mode: full
max_skills: 10
XML Output Format¶
When skills injection is enabled, agents receive structured XML in their system prompt:
<available-skills>
<skill id="python-style-guide" name="Python Style Guide" description="PEP 8 coding conventions">
<instructions>
## Python Style Guide
Follow PEP 8 conventions:
- Use 4 spaces for indentation
- Maximum line length of 88 characters
- Use snake_case for functions and variables
- Use PascalCase for classes
</instructions>
<base_directory>/home/user/.config/agentpool/skills/python-style-guide/</base_directory>
</skill>
<skill id="refactoring" name="Code Refactoring" description="Safe refactoring techniques">
<instructions>
## Code Refactoring
Always follow these steps:
1. Understand the existing code
2. Run tests before changes
3. Make small, focused changes
4. Run tests after each change
5. Commit incrementally
</instructions>
<base_directory>/home/user/.config/agentpool/skills/refactoring/</base_directory>
</skill>
</available-skills>
Complete Example¶
# Global skills configuration
skills:
paths:
- ~/.config/agentpool/skills
- ./project-skills
include_default: true
# Default: metadata-only injection for all agents
instruction:
mode: metadata
max_skills: 20
agents:
# Appends /skills to tool names (default: false)
append_tools_namespace: true
# Standard agent - uses global metadata injection
assistant:
model: openai:gpt-4o-mini
system_prompt: "You are a helpful assistant"
tools:
- type: skills
# Expert agent - uses full skill content
expert:
model: openai:gpt-4o
system_prompt: "You are an expert developer"
tools:
- type: skills
injection_mode: full
max_skills: 10
# Minimal agent - no skills injection
minimal:
model: openai:gpt-4o-mini
system_prompt: "Keep responses brief"
tools: [] # No skills tool
Backward Compatibility¶
By default, skills injection is disabled (mode: off). This ensures:
- Existing configurations continue to work unchanged
- Agents without explicit configuration see no skill injection
- Opt-in required to enable automatic injection
Related Configuration¶
- Toolsets - Configure agent tools including skills tool
- Agent Pool - Global pool configuration
See Also¶
- RFC-0008: Dynamic Skills Injection - Implementation details
- Skills Toolset - Skills toolset reference