Engineer Code

Use this skill to help users apply Microsoft Trace to optimize Python code implementations instead of treating Trace as a generic prompt-tuning tool.

Read references/microsoft-trace-trainable-methods.md before giving detailed guidance. It summarizes the core Trace workflow, code patterns, and caveats with citations to the Trace repository, docs, paper, examples, and generated API docs. When the task resembles an official Trace example pattern, map to the closest example before inventing a larger custom design.

When to use this skill

• The user wants to optimize Python behavior with Microsoft Trace or trace-opt.
• The user wants to convert a prompt, helper function, or small agent method into a trainable Trace component.
• The user asks how to use trace.node, @trace.bundle, @trace.model, ExecutionError, or OptoPrime.
• The user wants to train code from unit-test results, benchmark scores, natural-language critiques, compiler errors, or other non-scalar feedback.
• The user wants help deciding what should stay plain Python and what should become a Trace parameter.

Do not use this skill as the primary fix for generic Python runtime performance tuning, missing Trace installation, infrastructure or tooling failures, or tasks with no repeatable feedback signal. In those cases, say Trace is a bad fit or only a secondary lever.

Required inputs