Sometimes the Best Prompt Is /new
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
When an LLM chat session goes off track after multiple turns, it often can't be steered back on course because each new response is conditioned on everything already generated, including flawed prior interpretations and assumptions. This effect, described in research as 'Lost in Conversation,' causes multi-turn performance to drop 39% on average across large-scale simulated conversations. The recommended fix is to recognize when a conversation is lost and restart it with /new or /clear, or switch to a different agent or model, rather than trying to argue the model back to the right path. A complementary strategy is running the same task in parallel across multiple agents or models with varied few-shot prompts, trading extra tokens and cognitive load for faster, better-informed decisions. None of this removes the manual work needed to make generated code production-ready.
Questions this post answers
Why can't I get an AI coding agent back on track once the conversation goes in the wrong direction?
Each new response from a large language model is generated in the context of everything said before, including prior wrong interpretations, conclusions, and assumptions. Adding a 'new direction' prompt just adds to that context rather than replacing it, so the model still weighs the accumulated wrong context more heavily. This effect is called 'Lost in Conversation' and multi-turn performance drops 39% on average across more than 200,000 simulated conversations. daily.dev surfaces practical prompting techniques for developers steering AI coding agents day to day.
What should I do when an AI chat agent gets stuck giving the wrong kind of answers after many turns?
Restart the conversation using a command like /new or /clear, or switch to a different agent or model entirely, instead of trying to argue the model back onto the right path. Because each turn is conditioned on all prior turns, steering rarely works once a session is far enough along; a fresh context avoids the buildup of wrong interpretations and assumptions. Developers refining agent workflows can follow more prompting strategies like this on daily.dev.
How can I compare different AI coding approaches faster when working with agents and models?
Run the same task in parallel across multiple agent harnesses and models using few-shot prompting with varied directions, then compare the resulting implementations side by side. This costs extra tokens and cognitive load but produces several shapes of the same solution, speeding up the decision of which approach to pursue, though it does not remove the manual work of making code production-ready. daily.dev helps developers weighing multi-agent workflows keep up with emerging AI coding practices.
45 Impressions1 Comment