What problem does it solve? Building complex PyPTO kernels for Ascend NPU operators in one shot leads to hard-to-debug failures and precision errors. This Skill enforces a disciplined decomposition-and-construction workflow so each kernel is split into verifiable semantic modules before integration. ## Core Features & Use Cases - Decomposition Gate: Reads module_count from DESIGN.md §0.3 to choose the L0 single-module path or the L1 multi-module path, consuming architect-defined breakpoints without inventing new ones. - Module Boundary Rules (R1-R6): Enforces that every module contains at least one heavy op (matmul, cross-tile reduce, scan), merges light ops into neighboring heavy modules, and keeps pypto.view/pypto.assemble as entry/exit fixtures rather than standalone modules. - Module-at-a-time Construction: Builds staged files (_module1.py → _module12.py → …) where each iteration adds at most one real module, cross-checks the golden function inventory before running, and freezes verified modules. - Use Case: When implementing a fused BatchMatmulMaxSum operator in PyPTO, use this Skill to split the matmul, max-reduce, and sum-reduce stages into contract-defined modules, validate each boundary tensor against golden output, and freeze them before integration. ## Quick Start Ask the AI to decompose the current operator into PyPTO modules according to DESIGN.md and construct them one at a time with golden boundary validation.
