Target: turn clusters into segment profiles with one targeting action each. Playbook: pandas_master.
1. Cluster, then profile in business columns
k-selection via elbow + silhouette jointly (never one alone). Profile each cluster on spend, frequency, recency, tenure — not on scaled PCA axes nobody can act on.
2. Name segments for decisions
'High-value lapsing' beats 'Cluster 2': the name must imply the action (win-back offer vs upsell vs ignore). Every segment gets exactly one owner and one next action.
3. Stability check
Re-run with a different seed/subsample; segments that dissolve were noise. Report the stability result in the readout — unstable segments get no budget.
Recap
Elbow + silhouette for k, profiles in business columns, action-implying names, and a seed-stability check before any targeting spend.