The number of districts to select depends entirely on what population you want to generalize your results to. You should not select districts randomly because if you do, the district defines the cluster, and you will have fewer than 12 very large clusters. This will make your design effect very, very large, resulting in substantial imprecision in your overall estimate of the prevalence of whatever form of malnutrition you are interested in. The resulting level of imprecision will make your results useless to make program decisions, which of course is the reason you are doing a survey. For example, you may end up with an estimate of the prevalence of acute malnutrition of 11.4% with confidence intervals of 1.8 - 23.1%. An estimate like this is utterly useless and means you wasted all the resources and effort used to do the survey.
You should first decide what population you are interested in, then select 25-30 clusters from within this population without regard to districts boundaries. If you want district-specific estimates, you will have to apply a calculated sample size to each district.
However, in order to do appropriate sampling, you should consult some guidelines, or better yet, get the assistance of someone with experience in surveys and sampling. There are many more potential pitfalls in sampling and survey implementation than can be addressed in this short post.