reasoning-trace-optimizer

To debug an AI agent repeatedly failing to use a specific tool correctly, you can analyze its reasoning traces to pinpoint tool confusion. This skill captures detailed thinking blocks between tool interactions, identifying exactly why the tool is misunderstood and suggesting prompt fixes.

Context degradation and instruction drift are common agent failure patterns where the AI loses track of original goals over multiple steps. This skill detects these issues by analyzing interleaved thinking blocks recorded during tool interactions using MiniMax M2.1.

You can optimize prompts by feeding agent reasoning traces into an automated analyzer. This skill evaluates captured thinking blocks and tool interactions to detect failure patterns, automatically generating improved prompts to correct issues like instruction drift and tool confusion.

Yes, analyzing agent reasoning traces with this skill requires the Anthropic SDK installed in your environment. It also depends on pydantic, rich, and python-dotenv to process and display the interleaved thinking blocks effectively.

The best way to capture interleaved thinking blocks is using an automated trace analysis tool. This skill records detailed reasoning steps between tool calls natively, addressing context degradation by preserving the exact thinking process of MiniMax M2.1.