📄 中文摘要
在使用大型语言模型（LLM）进行多步骤任务时，常常会遇到输出不一致的问题，例如第一步表现良好，第二步却遗漏了关键约束，第三步自信地修改了不希望更改的内容，第四步与第一步相矛盾。这并不是模型的懒惰，而是因为在没有稳定参考的情况下，模型难以处理多个动态因素。为了解决这个问题，提出了一种称为“工作集”的简单提示模式。工作集是一个小而明确的上下文包，需保持更新并在每次开始新的子任务时重新发送或重申，类似于迷你项目的README、草稿和验收标准。
工作集提示：如何在多步骤工作中保持大型语言模型输出的一致性
📄 English Summary
The Working Set Prompt: How to Keep LLM Outputs Consistent Across Multi-Step Work
When using a large language model (LLM) for multi-step tasks, inconsistencies in outputs often arise, such as great performance in Step 1, forgetting key constraints in Step 2, confidently altering unwanted content in Step 3, and contradicting Step 1 in Step 4. This issue is not due to the model being 'lazy,' but rather its struggle to juggle multiple moving parts without a stable reference. To address this, a simple prompt pattern called the 'Working Set' is proposed. A Working Set is a small, explicit bundle of context that should be kept up to date and re-sent or re-asserted whenever starting a new sub-task, akin to a mini project README, scratchpad, and acceptance criteria.