What problem does it solve? Capturing a human expert's reusable GPU optimization recipe so GEAK's e2e_workflow and kernel_workflow can reproduce it automatically requires a strict contract: correct operator alignment, required documentation sections, and a two-sided on-box validation gate. This Skill walks a contributor through scaffolding, filling, validating, and submitting such an expert skill without missing any gate. ## Core Features & Use Cases - Scaffolding: Generates a new skill skeleton via scaffold.py with operator, scope (kernel or e2e), gens, dtypes, and regimes, rejecting operators not present in capability_index.yaml. - Two-sided validation: Runs static schema checks, emits the exact on-box measurement command for the matching workflow, and records measured efficacy (isolated speedup or e2e delta) plus do-no-harm control results before stamping a skill as validated. - PR submission: Branches, commits the skill and regenerated index.yaml, pushes, and opens a pull request, refusing unless the skill status is validated. - Use Case: An engineer who measured a 67% e2e gain from a FlyDSL fp8 blockscale down-proj playbook on gfx942 uses this Skill to turn that playbook into a validated, selector-matchable expert skill that future GEAK runs can auto-apply. ## Quick Start Ask the agent to scaffold a new expert skill for a given operator and scope, fill in the recipe sections, then validate it on-box and open a PR.
