Google shows prompt quality matters

Google’s Gemini story here is less about a flashy new model and more about something more awkward — a lot of “bad AI” behavior is really bad context management. An Android Police piece published May 3 says Gemini’s answers kept getting slower and less relevant the longer one conversation thread got. The fix was not some genius prompt hack. It was starting a new chat when the old one had become a junk drawer. That lines up pretty neatly with Google’s own prompt guidance, which has gotten more explicit lately about specificity, examples, structure, and keeping tasks focused. (androidpolice.com) ### What actually went wrong? The reported mistake was simple: using one Gemini thread for too many related-but-not-identical tasks. At first that feels smart — you keep all the context in one place. But once the thread gets long and mixed, the model starts dragging old details into new requests, even when those details no longer help. In the Android Police test, that showed up as laggier responses and answers that missed the point. (androidpolice.com) ### Why would a long thread hurt? Because context is not memory in the human sense. A model does not “know what matters” unless the prompt makes that clear. If you keep feeding it a giant transcript, the model has to weigh a lot of stale material against the new request. Sometimes it picks the wrong signal. Basically, the thread turns into a(androidpolice.com)t better. This is partly an inference from how these systems use prompt context, but it matches both the observed behavior and Google’s advice to keep prompts focused and concise. (androidpolice.com) ### Was the fix really just “start over”? Pretty much. The writer says moving to a new thread made Gemini faster and more on-topic right away. That sounds counterintuitive because people assume preserving context is always helpful. But context only helps when it is relevant context. Once a conversation mixes multiple subtopics, the carryover becomes contamination. A fresh thread acts like a reset button without needing the model itself to change. (androidpolice.com) ### How does this fit Google’s own guidance? Very closely. Google’s current Gemini prompt docs push clear and specific instructions, examples for in-context learning, delimiters that separate instructions from background, and shorter, more focused prompts. For Gemini Nano, Google is even more blunt: verbose preambles can hurt results, struct(androidpolice.com)roduct version of “don’t make one chat do everything.” (ai.google.dev) ### So is this really a prompt-engineering story? Yes — but not in the old “write a magic spell” sense. The interesting part is that prompt quality now includes thread hygiene, task boundaries, defaults, and examples. In January, Google also rolled out Automated Prompt Optimization for ML Kit’s GenAI Prompt API and said better system instruction can get quality close to fine-t(ai.google.dev)s a product surface, not just user text in a box. (android-developers.googleblog.com) ### Why does that matter for normal users? Because it changes where reliability comes from. People often wait for the next model release to fix bad outputs. But a lot of the gain may come from better scaffolding — clearer templates, narrower tasks, cleaner resets, better examples. In other words, the model can feel smarter without actually becoming smarter in the headline-grabbing way. (ai.google.dev) ### What should you do with Gemini now? Use separate chats for separate jobs. Put the actual task up front. Add constraints and examples when format matters. And if a thread starts feeling muddy, do not negotiate with it — start a new one. That is the small lesson here, but it points to a bigger one: AI quality is increasingly a UX problem. The winners may be the products that hide that complexity best. (androidpolice.com)