| name | mylab-structure-tuning |
| description | Use when the main loop is idea -> implementation -> training -> eval -> analysis for a structural experiment round. |
| trial_kind | structure-tuning |
Mylab Structure Tuning
Use this skill when the user is trying new model structures, new training logic, new loss wiring, or new module combinations.
Flow
- Idea
- Implement
- Train
- Eval
- Analyze
Frontmatter Essence
- current hypothesis or structural idea being tested
- exact implementation delta relative to the previous stable baseline
- expected learning signal from train plus eval plus analysis
- reason this round matters for the next design decision
- exact code checkpoint this round should start from
Trial Body Rules
- The execution trial should read like a causal chain from idea to analysis.
- The trial must make the implementation delta explicit.
- The trial must separate training from evaluation and analysis.
- The analysis step must say what kind of structural conclusion will be drawn.
- Deliverables should emphasize code diff, train/eval evidence, and design conclusions for this round only.
Reference Files
- references/trial-skill.md: the workflow contract that explains how this structure-tuning trial should be interpreted
- references/design.md: detailed hypothesis rationale and code-change logic
- references/experiment.md: detailed execution environment and artifact inventory
- references/analysis.md: detailed explanation of successes, failures, and anomalies
- references/conclusion.md: expanded verdict, limitations, and follow-up memory
- references/shared-asset.md: durable repository knowledge already learned in this run
- references/persistent-feedback.md: long-lived user guidance that should constrain future design choices
- references/recent-feedback.md: short-lived user feedback for this round
- references/parent-trial.md: previous trial body when iterating from an earlier trial
Templates