Lipid nanoparticle (LNP)-mRNA vaccines compressed the path from pathogen sequence to licensed product into less than a year and are now approved for two diseases, with more than twenty indications in clinical development. The platform's potency is no longer in doubt; its predictability is. We argue that progress is now limited by three interfaces that are engineered largely independently: the lipid-RNA interface, which sets encapsulation, internal structure and storage stability; the particle-cell interface, which sets tropism and the efficiency of endosomal escape; and the particle-immune interface, which sets the balance between adjuvanticity and reactogenicity. Quantitative design rules exist for the first interface, remain empirical for the second, and are almost absent for the third, which is why formulation development still proceeds by screening rather than by calculation. We review the molecular design of the mRNA payload, the chemistry, manufacture and intracellular trafficking of LNPs, the immunological chain that links innate activation to germinal-centre output, and the clinical evidence across COVID-19, respiratory syncytial virus, influenza, cytomegalovirus, HIV and other targets. We conclude by identifying the measurements and design rules that would make the next generation of LNP-mRNA vaccines predictable rather than empirical.