Research and pilot fabrication lines must plan capacity under highly diversified product portfolios, unstable demand, and frequent recipe changes. We present a prescriptive mixed integer linear programming framework that encodes internal mechanisms of cluster tools, including expected clean time after recipe transitions, finite concurrency at load ports and side buffers under a nonmixing policy, and sequential and parallel routing across chambers. The framework supports two planning tasks, minimizing added capacity to relieve bottlenecks and maximizing wafer starts for a given product mix and planning horizon. Using 12 monthly instances from an industrial R&D line, the proposed model improves agreement with the fab reported bottleneck set under substantial mix changes. Relative to a baseline formulation that abstracts cluster tool internals, the bottleneck hit rate increases by 0.0 to 10.5 percentage points across months (average improvement 4.8 percentage points). In addition, the model can screen incremental product releases that do not increase the modeled bottleneck loading, enabling approximately 1.2% additional starts without new equipment.