Opening the book…
A cipher's strength is not a property you can see by inspection; it is a residue left after many competent people have tried and failed to break it. Your own design has had one reviewer who already believes it works. When the specification leaks — and Kerckhoffs assumes it will — a homegrown algorithm loses everything at once, while a well-studied one loses nothing but its name.
Reach for the primitives that standards bodies and the open literature have converged on, and use them in the modes they were analysed in. When you need something unusual — a password hash, an authenticated stream, a threshold scheme — look for an existing construction before you compose one yourself. Treat any sentence beginning 'we XOR it with' as a design smell, and record which published construction each part of your system corresponds to.
Research and teaching exist precisely to produce new designs, and a novel scheme published for attack is doing the right thing. The rule bans deploying unreviewed cryptography to protect real data, not inventing it.