The core problem is that prompts grow like weeds. Engineers start with something simple, then keep adding fixes and new instructions without ever cleaning up what came before. Nobody reads the whole thing end-to-end, so contradictions pile up. You end up with thousands of lines where different sections tell the agent to do opposite things. The author, who leads AI engineering for OpenAI's startup team across Europe and Asia-Pacific, has seen this across companies with hundreds of millions in annual revenue—places with their own data annotation teams and custom models. Yet their prompts are full of logical errors that tank agent performance. The real kicker: after a few days rewriting prompts with these teams, he's seen 50% speed improvements, 40% retention gains, and 30% cost reductions. One Loveable engineer cut LLM spend by $20 million annually just by fixing inconsistencies his mother spotted. The second major failure is ambiguity hiding in plain language. When you write "never refer to competitors in output," it sounds clear to you because you live and breathe your company's competitive landscape. But the agent doesn't know who counts as a competitor, what about partial collaborators, or which situations apply. Engineers often read what they *meant* to write rather than what's actually there. Conditional prompts make this worse—different teams maintain different sections in separate files, so nobody ever sees the full picture and how pieces interact. The result is spaghetti: massive, tangled prompts where by the time you reach the end, you've forgotten the beginning or just stopped paying attention. The fix requires treating prompts like product decisions and code. Agent behavior shapes user experience as much as any UI does—whether it responds fast or thoroughly researches, whether it asks permission before acting, what language it uses. These are product choices, not just technical tweaks. So prompts need the same rigor as code: modular, logically consistent, actively maintained, and reviewed for contradictions. You can't expect an LLM to deliver a coherent experience if you can't specify what that experience actually is. The author proposes a structured template approach, though the article cuts off before revealing the full template itself.