Skill Discovery

Iron Law

NO TASK EXECUTION WITHOUT SKILL SEARCH FIRST.

Before starting any task, you MUST search for relevant skills. Even a 1% chance that a skill might apply means you should search.

When to Search

Decision Tree:
├── New task received?
│   └── YES → Search skills immediately
├── Encountering unfamiliar problem?
│   └── YES → Search skills
├── About to use a common pattern?
│   └── YES → Search for existing skill
├── Task similar to past experience?
│   └── YES → Search for induced skills
└── All else
    └── Search anyway (it costs nothing)

How to Search

1. •

   Use SkillSearchTool with a descriptive query:

   • •Include the problem domain: "memory leak debugging"
   • •Include the action type: "extracting entities from text"
   • •Include the context: "Neo4j graph optimization"
2. •

   Evaluate results using these criteria:

   • •Q-value > 0.7: High confidence, use directly
   • •Q-value 0.4-0.7: Moderate confidence, use with judgment
   • •Q-value < 0.4: Low confidence, verify before using
   • •No results: Proceed without skill, but log the gap
3. •

   Load matching skills with SkillLoadTool:

   • •Load the top 1-2 matching skills
   • •Read the full content before proceeding
   • •Follow the skill's instructions precisely

Anti-patterns

• •Skipping search because "this is a simple task"
• •Searching with vague queries ("help")
• •Ignoring search results because they seem familiar
• •Loading skills but not reading them
• •Using outdated mental model instead of checking for updates

After the Task

• •If a skill was helpful → provide "success" feedback
• •If a skill was NOT helpful → provide "failure" feedback
• •If no skill existed but should have → note this for skill creation