One text message from a phone can come back as a working site. Here is the assembly line that makes that possible. Click any station on the conveyor (or the buttons), and flip the prompt switch to see how input quality ripples through every stage.
Early code models emitted one blob of code and hoped. Modern agent pipelines loop:
The 3D conveyor above is honest about one thing: quality is decided mostly at the leftmost station.
| Failure | Cause | Fix |
|---|---|---|
| Beautiful but wrong content | Model invented facts you never supplied | Paste real copy, prices, names into the prompt |
| Works on desktop, broken on phone | No viewport constraint stated | Say "test at 380px width" explicitly |
| Placeholder images everywhere | Agent has no asset access | Provide URLs or accept generated SVG art |
| Half-finished features | Scope too big for one pass | Ship core first, iterate with follow-up prompts |
| Security holes in forms | Generated backend trusted user input | Ask for validation and review anything server-side |
Phone-to-production works because the pipeline is asynchronous:
| Stage | Time | Token cost share |
|---|---|---|
| Prompt intake + plan | 10–30 s | ~5% |
| Scaffold structure | 10–20 s | ~10% |
| Code generation | 1–5 min | ~50% |
| Test + self-repair loops | 1–10 min | ~30% |
| Deploy + preview URL | 10–60 s | ~5% |
Self-repair time is the wildcard: a vague prompt can triple it because the agent iterates against guesses instead of requirements.