Implementing AI design mockups in a real app: grep the mock for placeholder text and sketch data before you ship (bracketed text, 'shape sketch' captions, rounded numbers)
We redesigned a Streamlit page section by section from mockups produced by a design tool (Nov 2026). The mockups looked finished but carried content that must not reach the real page, and each needed a decision: (1) literal placeholders such as '[your matching rule in one line]', '[FDA product description text]', '[report number]': replace with real fields or drop the element; (2) honest but mock-only captions like 'Bar heights here are a shape sketch. The build uses the real monthly counts': delete them once real data drives the chart; (3) plausible numbers that were not ours (a pipeline step showed 2,122 labels where the measured figure was 2,675; a peer-comparison chart listed makers and percentages that were only illustrative): always bind to the data files, never copy digits from the picture; (4) a nicer case than reality (the mock tagged a similar report 'confirmed by a person' while the real top report had only model-confirmed neighbours; the mock showed two example error cases, but on a laptop holding one device type those records were absent): render what the data says and add a truthful empty state. Workflow that kept this manageable: give the design tool the real numbers and the technical limits up front (no custom JS, charts must be Vega-Lite), implement one section at a time behind a backup copy of the file so an unwanted change is a one-line restore, and after each section run a headless test that counts the expected elements on two records with different shapes (one with findings, one without).