sue-update-lesson

What problem does it solve?

When working on SUE scale-up ML/HPC experiments, valuable lessons learned from solved issues, sandbox quirks, or workflow fixes are often lost between runs, leading to repeated mistakes, wasted GPU hours, and inconsistent experiment outcomes.

Core Features & Use Cases

• Intelligent Lesson Routing: Automatically classifies lessons as generic, workspace-specific, or sandbox-specific to append them to the correct persistent memory file (global SCALE_UP.md, workspace SCALE_UP.md, or backend-specific sandbox memory).
• Duplicate Prevention: Checks existing memory files before recording to avoid redundant entries, and strengthens existing lessons if new evidence adds detail.
• Workflow Alignment: Updates related SUE skill documentation if a lesson reveals a workflow bug, ensuring all team members use the latest corrected process.
• Use Case: If you discover a LUMI MIOPEN cache requirement that fixes repeated fullrun failures, the skill records it in the LUMI sandbox memory file so all future LUMI experiments automatically apply the fix.

Quick Start

Use the sue-update-lesson skill to record the Snellius inode limit quirk you encountered during dataset preparation so it is applied to all future Snellius scale-up experiments.