The quality of compiler-optimized code for high-performance applications lags way behind what optimization and domain experts can achieve by hand. This paper explores in-between solutions, besides fully automatic and fully-manual code optimization. This work discusses how generative approaches can help the design and optimization of supercomputing applications. It outlines early results and research directions, using MetaOCaml for the design of a generative tool-box to design portable optimized code. We also identify some limitations of the MetaOCaml system. We finally advocate for an offshoring approach to bring high-level and safe metaprogramming to imperative languages.