How Do Anthropic Engineers Actually Prompt Claude Code?

TL;DR

Anthropic engineers prompt Claude Code through reusable skills rather than writing a new prompt for every task. Their approach follows four rules: prompt skills, build skills with descriptions, instructions, and tools, keep skills small and composable, and refine them after each session. Saved scripts and invocation controls can also make workflows more consistent and controlled. Read on for the practical patterns behind each rule.

Transcript

I listened to Anthropic's engineers at the AI Code Summit and I learned something I wasn't expecting. Almost everyone is prompting Claude code wrong. So, I decided to dig deeper and after studying everything Anthropic engineers have published, I uncovered four rules for how they actually prompt Claude code. And it turns out you don't need any techn... Read More

Key Insights

• Anthropic engineers focus on prompting skills, not individual prompts, to streamline AI interactions.
• Claude skills are organized collections of procedural knowledge, akin to apps on a smartphone.
• Skills consist of three layers: description, instructions, and tools, each playing a crucial role in task execution.
• Composable skills, rather than custom skills, allow for efficient task management and reusability.
• Scripts saved within skills enhance consistency and efficiency, reducing the need for repetitive prompt creation.
• User invocable and disable model invocation flags control who can access or run specific skills, enhancing security.
• Skills improve over time as users refine them based on past interactions, creating a compounding improvement loop.
• Regularly updating skills ensures they remain effective and adapt to changing user needs and AI capabilities.

Questions & Answers

Q: How do Anthropic engineers actually prompt Claude Code?

They focus on prompting reusable skills instead of creating a new standalone prompt for every task. The transcript identifies four rules: prompt skills, treat skills as more than prompts, build composable skills, and improve them after every session.

Q: What are Claude Code skills?

Skills are organized collections of files that package procedural knowledge for agents, meaning they capture a way to complete a task. For example, an email-response skill could be invoked with “/draft email” and supplied with the email that needs a reply.

Q: What are the three layers of a Claude skill?

The three layers are the description, instructions, and tools. The description helps Claude decide when to use the skill, the instructions provide the step-by-step playbook, and the tools can include code scripts, API calls, and reference files.

Q: Why should Claude skills be small and composable?

Small, focused skills can work together while remaining easier to diagnose, improve, and reuse. The speaker replaced one large content-creation skill with focused skills such as YouTube idea research, YouTube script writer, and LinkedIn post, which could then be chained together.

Q: Why save scripts inside Claude skills?

Saving a recurring script lets Claude rerun it in later sessions instead of rewriting it each time. The transcript describes saving a Python script used to style slides, making that work more consistent, efficient, faster, and repeatable.

Q: What do the user-invocable and disable-model-invocation flags do?

Setting user invocable to false hides a skill from the slash menu and reserves it for agents. Disable model invocation does the opposite: only the user can run the skill, which the transcript recommends for higher-risk actions such as sending a message or deploying code to production.

Q: How do Claude Code skills improve over time?

After each use, lessons about the user’s voice, process, rules, and edge cases can be written into the skill. Because the skill persists after the chat closes, the next session can begin with those improvements instead of starting over.

Q: When should code be used instead of AI inside a skill?

The transcript’s rule of thumb is to use code when code can perform the task. A script can be written once by AI and then reused, replacing repeated interpretation and token use with cheaper, faster, and repeatable code execution.

Summary & Key Takeaways

• Anthropic engineers prioritize prompting skills over individual prompts. Skills are organized knowledge collections that simplify task automation. This approach allows users to leverage AI without technical expertise, ensuring consistent task execution.

• Skills have three layers: description, instructions, and tools. Composable skills enhance task management and allow for easy updates and reusability. This method contrasts with traditional, custom skills that can become cumbersome.

• Saving scripts within skills ensures consistency and efficiency, while user invocable flags enhance security. Skills improve over time through regular updates, creating a feedback loop that enhances AI performance and user experience.