One scenario per test, ending in CLEAR_APP | A failure names one feature, and the next test starts from a known state |
| Copy the app's exact visible wording | Paraphrasing is the most common cause of a tap landing on the wrong element |
| Disambiguate, don't describe | Drizz picks one element out of everything on screen. A label, a position in a list, a neighbor or a section narrows it to one |
| Add a wait after anything that navigates or calls the network | Validations retry a few times, which covers a slightly slow screen, not a four-second one |
| Set a variable before comparing it | A number hardcoded last quarter fails when the price changes |
Validate directionally — greater than, less than | Survives price and tax changes; an exact computed total does not |
| Validate after anything non-deterministic | Scroll stops silently at its attempt limit, so the validation is what catches the miss. Applies after MAP_ACTION, after a load-bearing SET, and after a scroll |
| Put anywhere-popups in Memory | One rule replaces the same IF block in forty tests |
Put flow-specific popups in IF blocks | An org-wide rule makes every test in the organization pay for a check only one flow needs |
| Put anything repeated into a module | A login flow that changes twice a year exists in one place |
| Quote scroll targets and give a direction | An unquoted target with no direction doesn't resolve |
Use Tap whenever the element has a visible label | MAP_ACTION is the least deterministic command in Drizz |
Validate a variable's value — {{order_id}} | order_id without braces is searched for as literal on-screen text |
| Store only what's rendered on screen | Vision AI reads pixels. System timestamps and internal IDs aren't visible to it |
| Reset state explicitly on iOS — log out, or reinstall | CLEAR_APP is effectively a no-op on iOS |
Use Tap on the back button for cross-platform back | PRESS_DEVICE_BACK_BUTTON doesn't exist on iOS |
Pick one style per value — {{phone}} | <phone> is inert text. Drizz types the literal characters |
| Bind a dataset to any plan whose tests use variables | A plan with variables and no test data bound can misread SET steps |
Name variables for the moment — total_before | Reusing a name overwrites the first capture permanently |
| Put environment and account values in a dataset | One script covers staging and production |
| Work the healed-steps list as a maintenance queue | A step that heals every run has a bad description |
| Delete tests that no longer mean anything | A green suite nobody trusts is worse than a smaller one people do |