Domain-specific languages (DSLs) have emerged as a plausible way for non-programming experts to efficiently express their domain knowledge. Recent DSL research has taken a technical perspective on how and why to create DSLs, resulting in a wealth of innovative tools, frameworks and technical approaches. Less attention has been paid to the design process. Namely, how can it ensure that the created DSL realises the expected benefits? This paper seeks to answer this question when designing DSLs for highly specialised domains subject to resource constraints, an evolving application domain, and scarce user participation. We propose an iteration of alternating activities in a human-centred design method that seeks to minimise the need for expensive implementation and user involvement. The method moves from a low-validity exploration of highly diverse language designs towards a higher-validity exploration of more homogeneous designs. We give an in-depth case study of designing an actuarial DSL called MAL, or Management Action Language, which allows actuaries to model so-called future management actions in asset/liability projections in life insurance and pension companies. The proposed human-centred design method was synthesised from this case study, where we found it useful for iteratively identifying and removing usability problems during the design.