{"slug": "sometimes-the-best-prompt-is-new", "title": "Sometimes the Best Prompt Is /new", "summary": "A developer's guide to managing LLM conversations suggests that restarting a chat with /new or /clear is often the best prompt to fix a failing agent, citing the 'Lost in Conversation' effect (arXiv:2505.06120) where multi-turn context biases the model. The author recommends parallel implementations across different models and few-shot prompting to evaluate trade-offs faster, while noting that human back-and-forth remains necessary for production-ready code.", "body_md": "`/new`\n\nor `/clear`\n\nis sometimes the best “prompt” you can give to an agent to fix a change going bad.\n\nOnce a chat with an LLM gets far enough or has multiple turns of interaction, it can trigger an effect called [“Lost in Conversation”](https://arxiv.org/abs/2505.06120), and it will be very hard to drive it out of the chosen path.\n\nThere are multiple mechanisms for this, but one useful thing to remember is that at each turn the next token is conditioned on everything already generated.\n\nA multi-turn conversation can be represented like this:\n\n```\nanswer₁ = f(prompt₁)\n\nanswer₂ = f(prompt₁, answer₁, _prompt₂_)\n\nanswer₃ = f(prompt₁, answer₁, prompt₂, answer₂, _prompt₃_)\n```\n\nEach turn, the LLM does not consider your prompt independently; it considers it in the context of everything else that has been generated. That is the current context.\n\nThere is also in-context learning, where the model responds to statistical patterns in conversations that affect direction.\n\nWhen you go into a multi-turn conversation, you have something like:\n\nYou might think that in a multi-turn conversation, if you ask the LLM to change direction to an alternative, the interaction might be:\n\n```\nLLM + original question \n+ \nnew direction\n```\n\nBut actually what is happening looks more like this:\n\n```\nLLM\n +\noriginal question\n +\n20 previous interpretations\n +\n20 previous conclusions\n +\nimplicit assumptions\n +\nterminology established during the conversation\n+ \nnew direction\n```\n\nIt will be everything else + `new direction`\n\n, in the context of everything else, which has much more and stronger information about all the other direction(s) than the new one.\n\nThis is how transformers work.\n\nYour job is to recognize when you are lost in conversation and act on it. Restart the chat or shift to a different agent or model to get a fresh perspective on the same problem.\n\nAnother strategy to win time is to have multiple different agent harnesses and models implement the same task in parallel. You will pay extra tokens *AND* cognitive load, but you will get multiple shapes of the same implementation.\n\nSo your decision will be faster.\n\nOn my personal account, when I reach the limits for one agent/model, I create variations of the prompt and parallel implementations for each. To make this work, you have to make sure that you use few-shot prompting so you give different directions for building.\n\nNone of these bakeoffs or prompt-variant strategies eliminate the back-and-forth required to make code production-ready. That is still your job.\n\nWhat it does is give you more options to evaluate and a clearer view of the trade-offs in each approach.\n\nUnderstanding matters more than speed.", "url": "https://wpnews.pro/news/sometimes-the-best-prompt-is-new", "canonical_source": "https://allaboutcoding.ghinda.com/sometimes-the-best-prompt-is-new/", "published_at": "2026-08-26 09:34:44+00:00", "updated_at": "2026-08-26 14:13:49.918048+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-tools"], "entities": ["arXiv", "Lost in Conversation"], "alternates": {"html": "https://wpnews.pro/news/sometimes-the-best-prompt-is-new", "markdown": "https://wpnews.pro/news/sometimes-the-best-prompt-is-new.md", "text": "https://wpnews.pro/news/sometimes-the-best-prompt-is-new.txt", "jsonld": "https://wpnews.pro/news/sometimes-the-best-prompt-is-new.jsonld"}}