empirical-prompt-tuning - Empirically Tune Prompts with Fresh Agents
Iteratively improves agent instructions by testing fresh subagents across scenarios and evaluating both self-reported ambiguity and instruction-side metrics.
Tags
Updated: 2026-10-08Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Check description-body alignment
- Dispatch fresh subagents
- Run scenario evaluations
- Measure quality and usage
- Trace ambiguity phases
- Apply minimal prompt fixes
- Maintain failure pattern ledger
- Assess convergence and overfitting
Inputs
- Target prompt text or path
- Evaluation scenarios
- Requirements checklist
- Failure pattern ledger
Outputs
- Iteration evaluation report
- Revised prompt
- Quality and usage metrics
- Failure pattern ledger updates
- Convergence decision
Requirements
- Task tool access
- New subagent dispatch capability
- Agent usage metadata
- Separate fresh subagent per iteration