ROXy (@VK_ROXy) “17 checkboxes. 2 vehicles. 1 sketch. The order Claude reads them in decides who'” — TopicDigg

17 checkboxes. 2 vehicles. 1 sketch. The order Claude reads them in decides who's at fault. What 24 minutes of Anthropic actually buys you: 0:00 — Hanna and Christian open, neither speaks Swedish 0:52 — The Swedish insurance claim scenario, unpacked 2:01 — Naked prompt fails: Claude calls it a skiing accident on Schabongatan 3:51 — Anatomy of a prompt that survives production 5:31 — Task and tone context do the heavy lifting 8:43 — Static background data belongs in the system prompt 9:36 — XML tags, the structure Claude was fine-tuned on 12:50 — Few-shot examples for the gray-area cases 15:50 — The final task reminder that kills hallucinations 20:00 — Output formatting and hard constraints 22:11 — Prefilled responses put words in Claude's mouth 23:17 — Extended thinking as a prompt-engineering crutch The baseline at 2:01 doesn't just underperform. Same Claude 3.5 Sonnet at temperature 0, a one-line prompt, and the model decides it's a skiing accident on a Swedish street called Schabongatan. Ship that into a claims pipeline and every misrouted file costs an adjuster's hour to unwind. The turn comes at 15:50. Hanna forces read-order: form first, then sketch. Reading the 17 checkboxes first gives Claude the vocabulary to parse a stranger's scribble. That single instruction moves a hallucinating model to "vehicle B at fault," box 1 for vehicle A, box 12 for vehicle B. Sequence killed the hallucination. Not a bigger model. Not a longer prompt. Just the order. At 22:11 Christian prefills the response with an opening XML tag. No preamble. No apology. Just the structured verdict your SQL database can swallow. That's the whole gap between demo and production. Skip the 24 minutes and you keep paying a claims adjuster to clean up after a chatbot that guessed skiing.👇