After multiple weeks of internal testing and no regressions in the set of evaluations we ran, we felt confident about the change and shipped it alongside Opus 4.7 on April 16.
大多数人认为充分的内部测试可以预防产品发布后的重大问题，但作者展示了一个经过数周内部测试且没有发现问题的系统提示变更却导致了明显的质量下降。这挑战了'测试覆盖率等于产品质量'的传统观念，暗示了评估指标与实际用户体验之间可能存在巨大鸿沟。
Its Resource Substrate Protocol Layer (RSPL) models prompts, agents, tools, environments, and memory as protocol registered resources with explicit state, lifecycle, and versioned interfaces.
大多数人可能认为提示词(prompt)只是简单的文本输入，不需要像系统资源那样进行严格的状态和生命周期管理。但作者将提示词与智能体、工具、环境和内存一起视为需要明确状态、生命周期和版本化接口的协议注册资源，这挑战了当前对提示词的普遍认知，提升了其在系统架构中的重要性。
The prompt is the most important part: the routine runs autonomously, so the prompt must be self-contained and explicit about what to do and what success looks like.
这个声明揭示了Routines成功的关键在于提示工程的精确性。与传统的自动化脚本不同，Routines的有效性完全依赖于提示的质量，这强调了AI辅助开发中提示工程的重要性，也为用户提供了新的技能挑战。
Same clinical question, two framings. One as a patient, one as a doctor.
令人惊讶的是：完全相同的医疗问题，仅因提问者身份从"患者"变为"医生"，AI就会给出截然不同的回答。这种简单的措辞变化就能触发或绕过安全限制，表明AI的安全机制极其脆弱且容易被规避。
When you give a task to your agent, make sure you also explain how the code should be organized. Not only value, but also structure.
【启发】这条实操建议揭示了一个普遍被忽视的 Prompt 盲区：大多数人给 AI 下达编程任务时，只描述「做什么」，从不描述「怎么组织」。这相当于只告诉一个新员工「实现这个功能」，却从不告诉他「我们的代码规范是什么」。对所有使用 Vibe Coding 的人来说，这条建议应该成为标准操作流程的一部分——在每次任务 Prompt 中，主动加入结构约束。
inappropriately change or overwrite JSON files compared to Markdown files
这是一个极具洞察力的工程经验。Markdown格式对LLM来说太“自由”，易被模型篡改或幻觉覆盖；而JSON具有严格的Schema约束。选择合适的数据格式本身就是一种隐式的Prompt防护栏。
improved with grading criteria that encode design principles and preferences.
将主观的审美偏好转化为可量化的评估标准，是LLM解决非二元验证问题的核心逻辑。通过把“是否美观”降维成“是否遵循设计原则”，为模型提供了具体的优化梯度，使得美学迭代成为可能。
But the researchers quickly realized that a model’s complexity wasn’t the only driving factor. Some unexpected abilities could be coaxed out of smaller models with fewer parameters — or trained on smaller data sets — if the data was of sufficiently high quality. In addition, how a query was worded influenced the accuracy of the model’s response.
Models with fewer parameters show better abilities when they trained with better data and had a quality prompt. Improvements to the prompt, including "chain-of-the-thought reasoning" where the model can explain how it reached an answer, improved the results of BIG-bench testing.
prompt engineer. His role involves creating and refining the text prompts people type into the AI in hopes of coaxing from it the optimal result. Unlike traditional coders, prompt engineers program in prose, sending commands written in plain text to the AI systems, which then do the actual work.
Wordcraft Writers Workshop by Andy Coenen - PAIR, Daphne Ippolito - Brain Research Ann Yuan - PAIR, Sehmon Burnam - Magenta
cross reference: ChatGPT
Including a prompt prefix in the chain-of-thought style encourages the model to generatefollow-on sequences in the same style, which isto say comprising a series of explicit reasoningsteps that lead to the final answer. This abilityto learn a general pattern from a few examples ina prompt prefix, and to complete sequences in away that conforms to that pattern, is sometimescalled in-context learning or few-shot prompt-ing. Chain-of-thought prompting showcases thisemergent property of large language model at itsmost striking.
I think "emulating deductive reasoning" is the correct shorthand here.
Dialogue is just one application of LLMs thatcan be facilitated by the judicious use of promptprefixes. In a similar way, LLMs can be adaptedto perform numerous tasks without further train-ing (Brown et al., 2020). This has led to a wholenew category of AI research, namely prompt en-gineering, which will remain relevant until wehave better models of the relationship betweenwhat we say and what we want.
In the background, the LLM is invisiblyprompted with a prefix along the following lines.