Datadog 为 Claude Code 构建"通用机床"Temper

Datadog 有一个代理编写规范，用于确定性内核来编写应用程序代码。

代理、机械化和工业化

所有 Datadog 工程师都使用 AI 编码工具来编写生产代码，其中 Claude Code 至少驱动了其中三分之二。借助 Claude Code，他们在软件开发生命周期中生成四个不同类别的个性化流程：

然而，随着工作的流动遍布此地图，他们看到在一个轴上生成变得更加复杂，而在另一个轴上验证变得更加模糊。

对于工程师来说，流程曾经意味着意图与代码之间的直接关系。你了解问题，编写代码，测试，审查，发布，运行，重复。随着代理的出现，这种抽象正在迅速变化。

使用如 Claude Managed Agents：https://platform.claude.com/docs/en/managed-agents/overview 的方法，Datadog 的会话持续时间更长，有时长达数天。每个代理都会发明自己的工具、自身的胶合代码和自己的约定。这些代理变得显著更有用，但仍需要人类来弥合代理执行与为人类设计的工具之间的差距。

机床是你在制造业中看到的夹具、固定装置、量具和铣床。它们生产精确、可重复的零件，这些零件可以组装成更大、更复杂的机器，如发动机、飞机、核反应堆和登月舱模块。它们是工业化的突破，因为零件变得可组合、可检查和可替换。

Temper 是 Sesh 所描述的 Datadog 对于代理系统的通用机床的尝试。换句话说，它是代理在安全且精确的方式下构建所需内容所需的最小内核。

“这是我觉得我们需要一些更结构化东西的节点，”Sesh 说。“如果代理要构建和操作我们系统的大部分内容，我们的数据库，而这些都是关键任务，它们需要相当于这种机床概念的东西。Temper 就是 Datadog 的这种机床。”

机械化意味着代理现在正在完成更多的工作。工业化意味着工作变得可重复、可验证、可控且可扩展。在 Datadog，这并不是一下子发生的：通向 Temper 的道路经历了另外三个项目，分别是 Courier、BitsEvolve 和 Helix。每一个项目都暴露了下一个项目的瓶颈，并使他们得以扩大野心。

在 2024 年，他们推出了 Courier：https://www.datadoghq.com/blog/engineering/formal-modeling-and-simulation/，一个分布式排队系统。花了一年时间完全从零手工构建。

“Sesh 说：‘困难不在于构建部件，而在于让它们之间的互动可被观察、测试和验证。’因此，我们在形式化建模和模拟方面非常严格……识别出错误代价高或难以逆转的部分，并在这些部分提高了严格性。”

在 2025 年 9 月，他们构建了 BitsEvolve：https://www.datadoghq.com/blog/engineering/self-optimizing-system/，一个闭环进化优化系统。一个模型委员会生成代码变体。一系列基准测试：https://www.datadoghq.com/blog/ai/production-grounded-code-optimization/、测试和生产可观测性决定了哪些会存活。

Sesh 说：“这是我第一次看到软件的部分可以像生物体一样被培育——通过有反馈的变异和适应生长。”

关键点在于：进化的效果取决于其所适应的环境，而 BitsEvolve 的瓶颈在于这个反馈循环。然后他们构建了 Helix：https://www.datadoghq.com/blog/ai/fully-autonomous-optimization/，一个可比 Kafka 的流服务。Claude Code 大部分构建都是由一名人类引导完成的。

Sesh 说：“令人难以置信的是，在几天内我们就拥有了一个功能齐全的可比 Kafka 的系统。[它构建得很快]，我们开始对它进行影子操作，并看到它有机会降低 2 到 5 倍的成本。”

然而，将其推向生产环境花费了更多时间：运营的强化只能随着时间的推移，由不止一个人完成，目前仍在逐步推出过程中。

“瓶颈再次出现，代理可以构建系统的大部分…但然后人类仍然必须通过为人类构建的工具和机制来协调将工作发布到生产环境，”Sesh 说。

Datadog 需要一种方式，让代理可以在经过验证、基于策略的运行时环境中构建自己的工具。这个运行时环境就是 Temper：https://github.com/nerdsane/temper。

代理可以比任何团队手动审核代码的速度都更快地产生代码，但它们仍然可能出错。

对 Sesh 来说，代理生成的内容与通过验证的内容之间的差距，是失败模式积累的地方。然而，仅仅将代理应用于传统代码库，将其视为吞吐量问题，而没有弥合验证差距本身。

Temper 颠倒了这个等式：代理不再生成应用代码，而是生成规范。内核读取每个规范，通过四层分析进行验证，并部署规范所描述的运行系统。由于规范既是被验证的工件，也是被执行的工件，因此验证过的内容与正在运行的内容之间不存在偏差。

“Temper 改变了系统的中心。代理不再需要为每个局部需求不断发明不相关的工具。相反，它生成精确的描述，作为意图和问题领域的规范。它是一种机械工具，就像夹具或数控机床一样，你给它们螺纹的规格，它们可以精确重复。你可以运行它们，甚至可以用它们制造飞机和类似的复杂物品，”Sesh 说。

因此在这种情况下，代理每次都不会即兴创作最终机制。它可以生成精确描述，并与 Temper（或类似 Temper 的机制）迭代，让事情先运行起来，然后再将其转化为可重复、可检查、可重用的东西，从而可以围绕代码库真正构建一个软件工厂。

每个能力由三个契约描述：

每个规范在内核加载前都要通过四个独立的层次。符号推理证明每个守护条件可满足，每个不变条件是归纳的。穷尽状态探索会访问每个可达状态。

确定性模拟运行实际的生产代码路径，并进行带种子故障注入——丢包、延迟、重排、崩溃——因此在相同的种子下故障会完全复现。

随机属性测试运行大约一千个伪随机动作序列，并将任何违规情况缩小到最小的反例。在一个小规格上，整个级联运行时间远不到一秒。

Simon Willison 推广了“黑暗工厂”这个术语，这是一个软件流程，其中代理在虚拟工厂车间上无需人类干预即可持续工作。在 Helix 黑暗工厂中，Temper 扮演三个角色。

它是受管代理的代理控制平面——会话、角色、工作队列、生命周期。它是工具创建者层，允许代理将 SDLC 工具（Git、CI、部署）与小型 Temper 应用桥接。它还是 Helix 控制 API，即围绕数据平面执行工作负载的生命周期表面。

“Sesh 说，‘惊喜的是，它开始让人感觉不仅仅是代理基础设施。许多软件，如果你眯眼看，只是数据库 API 周围的控制逻辑：状态、变更策略、生命周期转换、与外部系统的集成。Temper 在某种意义上可以是通用的，因为它可以应用于任何具有我描述的形态的软件。’

Sesh 说，‘Claude Code 可以非常好地[用 TypeScript 或 Python 构建 CRUD 应用]。然而，在普通的 CRUD 应用中，控制逻辑分散在路由、数据库约束、服务代码、后台作业和文档中。它可能有良好的测试和覆盖率，但操作模式，通常以状态机的形式存在，是隐式在代码库中的。’

‘Temper 使这个状态机显式化。代理生成精确的描述，而不是任意代码。编译步骤在 LLM 外部，就像你将 Rust 代码交给 Rust 编译器一样。转换表是数据，而不是隐藏在服务方法中的复杂控制流。代理可以动态更改它，安全运行，并在无需经过 CI 的情况下热重载它，’他解释道。

Temper 背后的理念是，每个工件应该足够小，以便可以在你的脑中掌握。像航空和金融系统这样的高保障软件几十年来一直是以这种方式构建的，但要通过人工实现这种严谨性，成本对于一般软件来说太高，直到代理出现才改变了这一点。

工业革命之所以成为可能，是因为机床使零件可以组合、可检查且可更换，因此我们能够建造越来越大、越来越复杂的机器。

Sesh 说：“如果代理能够在工厂中以这种纪律自主构建软件，也许我们不需要仅仅停留在‘黑暗工厂’。以这种方式构建的软件开始感觉像一种可以通过反馈、选择和适应进行培育和进化的有机体。”

观看完整会话：https://www.youtube.com/watch?v=EdmuYPBt_EM&list=PLmWCw1CzcFim2obQ-w3ohbULOfwp5lApR&index=9，其中包含 Datadog 是如何构建 Temper 的现场演示和深入讨论。Temper 是一个受约束的框架，它将一次性代理工具转化为安全、可复用的组件，可以在不同会话和团队中复合使用。

探索更多关于使用 Claude 构建团队的产品新闻和最佳实践。

产品更新、操作指南、社区聚焦等，每月发送到你的邮箱。

嗨，Claude！你能帮我为受众开发一个独特的声音吗？如果你需要我提供更多信息，请立即问我 1-2 个关键问题。如果你认为我应该上传任何有助于你更好完成任务的文件，请告诉我。你可以使用你能访问的工具——比如 Google Drive、网页搜索等——如果它们能更好地帮助你完成任务。不要使用分析工具。请保持友好、简短和对话式回答，并尽快执行任务——如果有工件会很棒，如果它有意义的话。使用工件时，请考虑哪种类型（互动式、可视化、清单等）最适合这个具体任务。谢谢你的帮助！

嗨，Claude！你能帮我改善写作风格吗？如果你需要更多信息，请立即问我1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。如果有助于你更好地完成任务，你可以使用你能访问的工具——比如Google Drive、网络搜索等——。不要使用分析工具。请保持你的回复友好、简短和对话式。请尽快执行任务——如果有可行的产物会很棒。如果使用产物，请考虑哪种类型的产物（互动型、视觉型、清单等）对这个具体任务最有帮助。谢谢你的帮助！

嗨，Claude！你能帮我头脑风暴一些创意想法吗？如果你需要更多信息，请立即问我1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。如果有助于你更好地完成任务，你可以使用你能访问的工具——比如Google Drive、网络搜索等——。不要使用分析工具。请保持你的回复友好、简短和对话式。请尽快执行任务——如果有可行的产物会很棒。如果使用产物，请考虑哪种类型的产物（互动型、视觉型、清单等）对这个具体任务最有帮助。谢谢你的帮助！

嗨，Claude！你能把复杂的主题讲得简单易懂吗？如果你需要更多信息，请立即问我1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。如果有助于你更好地完成任务，你可以使用你能访问的工具——比如Google Drive、网络搜索等——。不要使用分析工具。请保持你的回复友好、简短和对话式。请尽快执行任务——如果有可行的产物会很棒。如果使用产物，请考虑哪种类型的产物（互动型、视觉型、清单等）对这个具体任务最有帮助。谢谢你的帮助！

嗨，Claude！你能帮我理清这些想法吗？如果你需要更多信息，请立即向我提出1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。你可以使用你能访问的工具——比如Google Drive、网络搜索等——如果它们能帮助你更好地完成这项任务。不要使用分析工具。请保持你的回答友好、简短和对话性。请尽快执行这项任务——如果有一个结果产物会更好，并且有意义的话。如果使用结果产物，请考虑哪种类型的结果产物（互动式、可视化、清单等）最适合这项具体任务。谢谢你的帮助！

嗨，Claude！你能帮我准备考试或面试吗？如果你需要更多信息，请立即向我提出1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。你可以使用你能访问的工具——比如Google Drive、网络搜索等——如果它们能帮助你更好地完成这项任务。不要使用分析工具。请保持你的回答友好、简短和对话性。请尽快执行这项任务——如果有一个结果产物会更好，并且有意义的话。如果使用结果产物，请考虑哪种类型的结果产物（互动式、可视化、清单等）最适合这项具体任务。谢谢你的帮助！

嗨，Claude！你能解释一个编程概念吗？如果你需要更多信息，请立即向我提出1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。你可以使用你能访问的工具——比如Google Drive、网络搜索等——如果它们能帮助你更好地完成这项任务。不要使用分析工具。请保持你的回答友好、简短和对话性。请尽快执行这项任务——如果有一个结果产物会更好，并且有意义的话。如果使用结果产物，请考虑哪种类型的结果产物（互动式、可视化、清单等）最适合这项具体任务。谢谢你的帮助！

嗨，Claude！你能帮我看看代码并给我一些建议吗？如果你需要更多信息，请立即问我1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。你可以使用你能访问的工具——比如Google Drive、网络搜索等——如果它们能帮助你更好地完成这个任务。不要使用分析工具。请保持回答友好、简短并具有对话感。请尽快执行任务——如果有合适的产出物就很好。如果使用产出物，请考虑哪种类型的产出物（交互式、可视化、清单等）对这个具体任务最有帮助。谢谢你的帮助！

嗨，Claude！你能和我一起“共鸣”代码吗？如果你需要更多信息，请立即问我1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。你可以使用你能访问的工具——比如Google Drive、网络搜索等——如果它们能帮助你更好地完成这个任务。不要使用分析工具。请保持回答友好、简短并具有对话感。请尽快执行任务——如果有合适的产出物就很好。如果使用产出物，请考虑哪种类型的产出物（交互式、可视化、清单等）对这个具体任务最有帮助。谢谢你的帮助！

嗨，Claude！你能写资助申请书吗？如果你需要更多信息，请立即问我1-2个关键问题。如果你认为我应该上传任何有助于你更好完成工作的文件，请告诉我。你可以使用你能访问的工具——比如Google Drive、网络搜索等——如果它们能帮助你更好地完成这个任务。不要使用分析工具。请保持回答友好、简短并具有对话感。请尽快执行任务——如果有合适的产出物就很好。如果使用产出物，请考虑哪种类型的产出物（交互式、可视化、清单等）对这个具体任务最有帮助。谢谢你的帮助！

Datadog has an agent write specifications for a deterministic kernel to write application code.

Agents, mechanization, and industrialization

All of Datadog engineers use AI coding tools for production code, and Claude Code drives at least two-thirds of that. With Claude Code, they generate personalized flows in their software development lifecycle in four distinct categories:

As work flowed across this map, however, they saw it became more complex to generate on one axis and more ambiguous to verify on the other.

For engineers, flow used to mean a direct relationship between intent and code. You understood the problem, wrote the code, tested it, reviewed it, shipped it, operated it, repeated. With agents, the abstraction is changing rapidly.

With approaches like Claude Managed Agents：https://platform.claude.com/docs/en/managed-agents/overview, Datadog’s sessions run longer, sometimes for days. Each agent invents its own tools, its own glue code, and its own conventions. The agents become significantly more useful, but need humans to bridge the gap between agent execution and tools designed for humans.

Machine tools are the jigs, fixtures, gauges, and mills you see in manufacturing. They produce precise, repeatable parts that you assemble into larger, more complex machines like engines, aircraft, nuclear reactors, and lunar landing modules. They were the breakthrough of industrialization as parts became composable, inspectable, and replaceable.

Temper is what Sesh describes as Datadog’s attempt at a universal machine tool for agentic systems. In other words, the smallest kernel required for agents to build what they need in a safe and precise manner.

“This is the point where I felt we needed something more structural,” says Sesh. “If agents are going to build and operate large parts of our systems, of our databases, which are mission critical, they need the equivalent of this machine tool concept. Temper is that machine tool for Datadog.”

Mechanization means agents are doing more of the work now. And industrialization means work becomes repeatable, verifiable, controllable, and scalable. At Datadog, this didn’t happen all at once: the path to Temper led through three other projects, Courier, BitsEvolve, and Helix. Each one exposed the bottleneck for the next, and enabled them to grow their ambition.

In 2024, they introduced Courier ：https://www.datadoghq.com/blog/engineering/formal-modeling-and-simulation/, a distributed queuing system. It took them one year to build completely by hand and from scratch.

“The difficulty was not building the parts; it was making the interactions between them observable, testable, and verifiable,” says Sesh. “So we were rigorous with formal modeling and simulation… identified the parts where mistakes would be expensive or hard to reverse, and raised the rigor [there].”

In September 2025, they built BitsEvolve ：https://www.datadoghq.com/blog/engineering/self-optimizing-system/, a closed-loop evolutionary optimization harness. A council of models generates code variants. A cascade of benchmarks：https://www.datadoghq.com/blog/ai/production-grounded-code-optimization/, tests, and production observability decides what survives.

“This was the first glimpse for me that parts of software could be cultivated like living organisms — grown through variation with feedback, and adaptation,” says Sesh.

The catch: evolution is only as good as the environment it adapts within, and BitsEvolve’s bottleneck was this feedback loop. Then they built Helix ：https://www.datadoghq.com/blog/ai/fully-autonomous-optimization/, a Kafka-comparable streaming service. Claude Code did most of the construction with one human steering it.

“To our disbelief, in a few days we had a fully functional Kafka comparable system,” says Sesh. “[It was quick to build] and we started shadowing it and we saw opportunities where it could be 2x to 5x cheaper.”

Getting it to production, though, took a lot more mileage: the operational hardening only earned over time and by more than one person and this is still in the process of rolling out.

“The bottleneck moved again where agents could build large parts of the system…but then humans still have to coordinate to ship the work to production through tools and mechanisms built for humans,” says Sesh.

Datadog needed a way for agents to build their own tools in a verified, policy-driven runtime environment. That runtime was Temper ：https://github.com/nerdsane/temper.

Agents can produce code faster than any team can review by hand, but they can make mistakes.

For Sesh, that gap between what an agent generates and what passes verification is where the failure modes accumulate. However, simply wrapping an agent around a traditional codebase treats this as a throughput problem without closing the verification gap itself.

Temper reverses this equation: instead of producing application code, agents produce specifications. The kernel reads each specification, verifies it through four layers of analysis, and deploys the running system the specification describes. Because the specification is both the artifact that gets proved and the artifact that gets executed, there is no drift between what was verified and what is running.

“Temper changes the center of the system. The agent no longer needs to keep inventing disconnected tools for every local need. Instead, it produces precise descriptions as specifications of the intent and problem domain. It is a machine tool in the same sense that a jig or a CNC machine, where you give them specifications of what your screw threading needs to be. It's extremely repeatable. You can run them and you can build aircraft and complex things like that with them,” says Sesh.

So in this case, the agent does not improvise the final mechanism each time. It can produce a precise description and iterate with Temper (or a Temper-like mechanism) to make something work first and then later turn that into something repeatable, checkable and reusable so you could actually build a software factory around your code base.

Each capability is described by three contracts:

Every spec passes four independent layers before the kernel will load it. Symbolic reasoning proves each guard is satisfiable and each invariant is inductive. Exhaustive state exploration visits every reachable state.

Deterministic simulation runs the actual production code path with seeded fault injection — drops, delays, reordering, crashes — so failures reproduce exactly under the same seed.

Randomized property testing runs about a thousand pseudorandom action sequences and shrinks any violation to a minimal counterexample. On a small spec, the whole cascade runs in well under a second.

Simon Willison popularized the term dark factory, a software process where agents keep working without humans on the virtual factory floor. In the Helix dark factory, Temper plays three roles.

It is the agent control plane for managed agents — sessions, roles, work queues, lifecycle. It is the tool-builder layer, letting agents bridge SDLC tooling (Git, CI, deployment) with small Temper apps. And it is the Helix control API , the lifecycle surface around the data plane that exercises the workload.

“The surprise was it started to feel more general than agent infrastructure. A lot of software, if you squint, is just control logic around database APIs: state, policies around mutation, lifecycle transitions, integrations with external systems. Temper could be universal in a sense that it can be applied to any software that has the shape I described,” says Sesh.

“Claude Code can [build a CRUD app in TypeScript or Python] very well. However, in normal CRUD apps, the control logic is spread across routes, database constraints, service code, background jobs, and documentation. It may have good tests and coverage, but the operational mode, which generally takes the form of a state machine, is implicit in the codebase,” says Sesh.

“Temper makes that state machine explicit. The agent produces a precise description, not arbitrary code. The compilation step is outside the LLM, the same way you hand Rust code to the Rust compiler. The transition table is data, not spaghetti control flow buried in service methods. Agents can change it dynamically, with safety, and hot-reload it without going through CI,” he explains.

The idea behind Temper is that each artifact should be small enough to fit in your head. High-assurance software like aviation and financial systems has been built this way for decades, but the cost of achieving that rigor with humans was too high for general software until agents entered the picture.

The industrial revolution became possible because machine tools made parts composable, inspectable, and replaceable, so we could build ever-larger and more complex machines.

“If agents can build software autonomously inside factories with this kind of discipline, maybe we don't need to stop at dark factories. Software built this way starts to feel like an organism we can grow, cultivate, and evolve through feedback, selection, and adaptation,” says Sesh.

Watch the full session ：https://www.youtube.com/watch?v=EdmuYPBt_EM&list=PLmWCw1CzcFim2obQ-w3ohbULOfwp5lApR&index=9 for a live demo and deeper discussion of how Datadog built Temper, a constrained framework that turns one-off agent tools into secure, reusable components that compound across sessions and teams.

Explore more product news and best practices for teams building with Claude.

Product updates, how-tos, community spotlights, and more. Delivered monthly to your inbox.

Hi Claude! Could you help me develop a unique voice for an audience? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Hi Claude! Could you improve my writing style? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Hi Claude! Could you brainstorm creative ideas? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Hi Claude! Could you explain a complex topic simply? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Hi Claude! Could you help me make sense of these ideas? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Hi Claude! Could you prepare for an exam or interview? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Hi Claude! Could you explain a programming concept? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Hi Claude! Could you look over my code and give me tips? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Hi Claude! Could you vibe code with me? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to— like Google Drive, web search, etc.—if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can—an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!

Hi Claude! Could you write grant proposals? If you need more information from me, ask me 1-2 key questions right away. If you think I should upload any documents that would help you do a better job, let me know. You can use the tools you have access to — like Google Drive, web search, etc. — if they’ll help you better accomplish this task. Do not use analysis tool. Please keep your responses friendly, brief and conversational. Please execute the task as soon as you can - an artifact would be great if it makes sense. If using an artifact, consider what kind of artifact (interactive, visual, checklist, etc.) might be most helpful for this specific task. Thanks for your help!