Kelly is the academic answer; half-Kelly is the practical one. Full Kelly maximizes long-run growth but expects you to estimate edge precisely. Nobody does. Half-Kelly cuts variance roughly in quarters with only a small hit to growth — and forgives the error in your edge estimate that you cannot remove.
The two destructive habits: flat-staking on every pick regardless of edge, and freerolling on a streak. Flat-staking treats a 3% edge and a 0.5% edge identically. Freerolling spends an opening run before the model has even sampled enough trials to know whether you have an edge at all. Discipline is small. The damage of skipping it is not.