Skill Tuning

What it does

Skill Tuning is designed to automatically diagnose and optimize AI agent execution, addressing common problems that impact performance and output quality. It identifies issues like excessive token usage, fragmented state management, or coordination failures between different parts of an agent's workflow. The skill then applies targeted fixes based on pre-defined strategies, aiming for improved accuracy, relevance, and overall efficiency – all without direct user intervention beyond the initial request.

When to use it

• When you notice your AI agents are exhibiting inconsistent behavior or producing unexpected results.
• If an agent's performance is slow due to excessive token usage or inefficient data handling.
• To resolve issues related to fragmented state, unreliable chains of actions, or problems with error handling within the agent’s workflow.
• When you suspect underlying architectural or configuration problems are hindering your AI agent's effectiveness.

Key capabilities

• Autonomous Diagnosis: Identifies and categorizes execution issues based on pre-defined patterns and severity levels (P0 - P5).
• Problem Taxonomy & Tuning Strategies: Leverages detailed specifications (problem-taxonomy.md and tuning-strategies.md) to guide diagnosis and remediation.
• Gemini CLI Integration: Utilizes the Gemini CLI for deep analysis of complex issues.
• State-Driven Orchestration: Manages a workflow with distinct phases: initialization, requirement analysis, diagnosis (multiple types), report generation, fix proposal, application, and completion.
• Fix & Verification: Applies proposed fixes and automatically retests to ensure effectiveness.

Example prompts

• "Skill Tuning: Analyze my agent's performance."
• "Skill Tuning: Diagnose issues with token consumption in my skill."
• "Skill Tuning: I’m seeing inconsistent results; can you identify the problem?"

Tips & gotchas

• The skill requires access to and understanding of problem-taxonomy.md and tuning-strategies.md for effective operation.
• While automated, Skill Tuning focuses on specific categories of problems outlined in its specifications. It may not address all possible issues.
• The "Fix & Verify" process automatically applies changes; review the final report (state.final_report) to understand what adjustments were made.