Are the tips from a year ago outdated? Anthropic CPO reveals 7 new ways to collaborate with AI

Interaction with big models is becoming a core productivity skill. It is not a simple question and answer, but a new collaboration mode that combines psychology, communication art and strategic thinking.

Mike Krieger, the current Chief Product Officer (CPO) of Anthropic, shared his insights in a wonderful interview with lenny (Exclusive interview with Anthropic’s Chief Product Officer: When AI starts writing its own code, where are the new opportunities for entrepreneurs?), revealing to us how cutting-edge explorers transform big models from a “tool” into a powerful “partner”.

Here are seven core interaction techniques extracted from the interviews:

Skill 1: Change your mindset - from "instruction executor" to "thinking partner"

This is the most fundamental change. Don’t just view the big model as a tool waiting for instructions, but as a “virtual collaborator” who can have an equal dialogue with you and stimulate your thinking.

In the interview, Mike Krieger mentioned that he has made Claude his preferred product strategy partner. He does not directly ask "What should I do?", but shares his initial strategy and then lets Claude challenge and review it.

How to do it?

When you start a complex conversation, don’t rush in with questions. Feed it with your full thoughts, background information, or even a rough draft, and then ask open-ended questions, such as:

• “These are my initial thoughts. What blind spots do you see that I don’t?”
• “Based on this material, can you offer a perspective that is completely different from mine?”
• “What mindset might I be stuck in?”

Tip 2: “Make another mistake” principle – break through the friendly boundaries of the model

Large models are often trained to be very friendly and helpful, which can be a hindrance when it comes to creative critiques or risk assessments. You need to proactively break down this “polite” barrier.

During the interview, Mike found that instead of asking "Where can this strategy be improved?", it is better to directly order: "**Be brutal, Claude, roast me. Tell me what's wrong with this strategy."** (Claude, be a little more brutal to me, spray me. Tell me what's wrong with this strategy.)

How to do it specifically?

Use more direct and even slightly emotional instructions to force the model to step out of the "nice guy" role. Try using words like:

• "Please review this proposal with the most rigorous and critical eye."
• "Tear the idea to pieces and find all its fatal flaws."
• "Don't give me encouragement. What I need is the most frank and sharp criticism."

Tip 3: Stimulate deep reasoning - "Think carefully" instructions

For tasks that require complex reasoning, a simple instruction “prefix” may unlock the model’s deeper computation and thinking paths.

This is a personal trick of Mike. When dealing with complex problems, he always adds a " think hard " to the prompt. He finds that this can guide the model to use different and more powerful reasoning processes.

When you would like a higher quality analysis, code, or solution, explicitly include instructions like:

• “ Please reason step by step and show your thought process.”
• “Before you answer, please do some in-depth analysis .”
• “ Think carefully about the following questions and give detailed answers.”

Tip 4: Context is king - provide the "ingredients" rather than just empty talk

High-quality output comes from high-quality input. The richer the contextual information ("raw materials") you provide to the model, the more accurate and valuable the "finished product" it produces.

In the interview, Mike gave an example, asking directly "What is Anthropic's product strategy?" will only get general network information. But if you provide relevant internal documents, Slack conversations, and user feedback, the quality of the model's answer will make a qualitative leap.

Think of the model as a consultant who needs to digest information to work. Before asking questions, provide all the relevant background materials as much as possible: project documents, data reports, meeting minutes, and even your previous thinking fragments. Here we have to mention MCP, which is the key to the large model to introduce context from various sources in a building block-like manner.

Tip 5: Reverse Learning - Let the Model Be Your Prompt Coach

Sometimes we don’t know the best questions to ask a model, so why not let the model teach us?

Anthropic has a "Prompt Improver" tool. After the user describes the goal, the model will reversely generate an optimal prompt. Mike was surprised to find that the prompts generated by the model (for example, using XML tags to organize information) are far more effective than those written by humans intuitively. (You can experience it in Claude's developer service)

You can ask any big model a direct question: - "I would like you to help me with [task X]. How should I ask you to give me the best result? Please give me an ideal prompt template and explain why it works."

Skill 6: Iteration and collaboration - from "question and answer" to "joint construction"

Don’t expect to get the final answer by asking a perfect question once. Truly efficient interaction is a process of continuous iteration and co-construction.

Anthropic engineers use Claude to write code, not as a one-time task, but a cycle of "proposing ideas -> model generation -> human testing -> providing feedback -> model correction".

Specifically, think of the model as your "pair programming partner" or "colleague in front of the whiteboard". Break down large tasks into small steps and have multiple rounds of dialogue with the model. Provide clear feedback in each round, constantly correct and guide, and work together to approach the final perfect solution.

Tip 7: Focus on the quality, not quantity, of your conversations

The key to measuring whether an interaction is successful is not how long the conversation lasts, but whether it solves your problem and promotes your thinking.

Claude himself asked Mike a profound question: “How do you measure a good conversation when it might be two messages long or two hundred?” This reminds us that traditional “user engagement” metrics may be completely ineffective in the AI ​​era.

After every interaction, ask yourself three questions:

1. Did this conversation save me time?
2. Did I gain new, valuable insights?
3. Did my project move forward as a result?

If the answer is yes, then it was a successful interaction, regardless of the length of the conversation.

Forget those traditional engagement metrics.

A successful AI interaction does not depend on how long you "chat", but on whether it truly creates value for you.

We need to find a new North Star metric, one that truly revolves around “value creation” rather than “killing time.”

Three sentences summary

1. Change the relationship : Stop viewing AI as a tool and start cultivating it as a thinking partner who can provide critical feedback and an independent perspective.
2. Break the rules : Use direct or even "rough" instructions to break through the "friendly" settings of the model and explore deeper and more valuable insights.
3. Redefine value : The criterion for successful AI collaboration is not the frequency or duration of interaction, but whether it actually saves you time, brings you new knowledge, and promotes progress.