Evaluating Tracks

How to measure and improve the quality of your tracks.

Evaluation criteria

A good track should score well on these dimensions:

Dimension	What to measure	Target
Activation accuracy	How often the track activates for relevant tasks	>90%
False positive rate	How often it activates for irrelevant tasks	<5%
Instruction following	How well the agent follows the instructions	>85%
Output quality	Quality of the final output	Consistent

Testing methodology

1. Create test cases: Prepare a set of tasks that should and shouldn't activate the track
2. Run activation tests: Verify the track activates only for matching tasks
3. Run execution tests: Verify the agent follows instructions correctly
4. Review outputs: Check that outputs meet quality standards

Common issues

• Too broad description: Track activates too often — narrow the description
• Too narrow description: Track rarely activates — broaden the description
• Vague instructions: Agent doesn't follow steps — add more specificity
• Missing examples: Agent produces inconsistent output — add examples