Your First Manual Run Log
Explore how to create a manual run log that captures inputs, triage JSON outputs, and routing decisions in a structured table. Learn to identify routing errors clearly and maintain JSON consistency to provide observable artifacts for debugging and improving AI workflows.
We'll cover the following...
A standard chat transcript can obscure workflow failures because the conversation may remain coherent even when routing or decision logic is incorrect. To evaluate the workflow, use a manual run log: a simple table that records the input, triage JSON, and routing outcome side by side. By recording these elements in a structured format, a routing error becomes directly visible in the recorded fields rather than remaining implicit in the conversation. The key test is inspectability: another reviewer should be able to read one row and explain why the workflow took a specific route without reconstructing the full conversation.
To start, use a minimal column set and keep it consistent across rows: