In the process of communicating and collaborating with AI, the author has experienced repeated twists and turns and felt deeply inefficient. To this end, the author carefully created an “AI Collaborative Coach” prompt, which significantly improved the efficiency of collaboration with AI. This article will share the creation process and usage experience of this prompt to help readers improve the efficiency of AI collaboration.
In the past week, I have been busy using Cursor to reconstruct the architecture of the “prompt management assistant” to prepare for the next version of the major upgrade, allowing the AI to write more than N code and step on countless pitfalls.
I tried to write an agent logic to refactor the code warehouse, and found that the AI wrote a lot of tall code, but it was completely useless. That can only be changed back to manual thinking, using prompts to separate the data of each main logic of the code little by little.
Using manual thinking does solve the reconstruction of the code warehouse, but the disadvantage is that communicating with AI is too tiring, and you tell it to go east, but it wants to go west.
I said to Claude-4-opus: This place doesn’t need the purple bar at the bottom, understand?
As a result, it abruptly changed the interaction of the copy button to only purple, and the area I marked did not change at all.
At that moment, I really wondered if there was a problem with my ability to express myself or with the AI’s ability to understand.
I went to hold a small AI programming activity offline on Saturday, and saw that everyone actually had similar problems, obviously just wanted the AI to draw a simple front-end page, but the AI framed a React architecture.
Humans and AI are nominally cooperating, but in fact they are fighting separately.
So why is this reason?
To achieve these three challenges, product managers will only continue to appreciate
Good product managers are very scarce, and product managers who understand users, business, and data are still in demand when they go out of the Internet. On the contrary, if you only do simple communication, inefficient execution, and shallow thinking, I am afraid that you will not be able to go through the torrent of the next 3-5 years.
View details >
Because the contextual information output by humans to AI is not enough, AI can only guess human needs in many cases.
Let’s not think about talking to AI for a while, let’s first use job hunting to understand the importance of contextual information.
Now we are a person who is asking for a job as an AI product manager, we are going to a good friend to discuss the job search plan, everyone thinks the following two expressions, which is easier to achieve good discussion results:
Expression A
I am a product manager who has been working for 7 years and I want to find a job.
Expression B
I am a product manager who has worked for 7 years, I have x years of experience as an AI product manager, and I have done AI projects in xx industries.
The next direction I want to work in is XX position, the expected salary is xxx, and the acceptable place is xxx.
I hope the company’s commute time is 9 am and 6 pm.
Obviously, it is the B expression, and it is easier to discuss high-quality results. If A, everyone can’t grasp the key points after chatting for a long time.
When communicating with people, the richer the contextual information, the better the effect, and the same is true for communicating with AI.
But giving AI enough contextual information is actually a difficult thing, taking my programming scenario as an example, I have two major pain points:
1. There are some things I can’t understand, such as the purple bar, I can talk about this dimension at most
2. My programming skills are not that deep, and I can’t speak many professional terms
These two points together lead to a small detail interaction, often I have to change it for 1-2 hours at a time, which is really too inefficient.
So I’ve been wondering if there’s any way to make this a little easier.
I tried many ways, such as building more documentation in the code warehouse, such as telling it everything I was thinking when describing to the AI, which was a bit effective anyway, but the problem of inefficient communication with the AI still exists.
It wasn’t until I was pondering and thinking about my partner’s conversation on the subway that a voice suddenly skipped through my mind: AI is the best way to supplement the context.
I just need to tell the AI what is on my mind and let it discuss it with me for a while, and an accurate description with enough contextual information will come out!
Previously, I had a direct conversation with AI, but now it has become me, AI intermediary, AI dialogue, so I quickly created an AI intermediary prompt to test this logic, and found that the effect is really much better than before.
I will continue to use my programming case to give you examples:
This is my latest version of the “Prompt Management Assistant” page, but there are many problems in the prompt area: the current interaction of the copy button is not good, and there is no layout for modifying and deleting buttons.
I have asked the AI to change this page many times, I really can’t get the style effect I want, every time I adjust the details with the purple bar, the AI doesn’t know what results will shock me.
So I asked the question to the “AI Collaborative Coach” prompt, and it started asking me various questions to gather information.
After gathering information and discussing, the “AI Collaborative Coach” gave me a command to get Claude4opus to work:
Then the AI gave me 3 copies of such a design drawing, its size and browser plug-in corresponded, so I could refer to that interaction I liked the most.
In the end, I chose option 1, but I didn’t know what to do in the case of various long texts, so I pulled the “AI collaborative coach” to chat for a few rounds, and finally we let Claude4opus produce a fairly satisfactory style rendering.
Next is the most troublesome part, how to move the style on the renderings to the code of the plugin, I had to adjust it for half a day every time I did this link before.
Because it’s useless if you just take a screenshot and tell the AI that I want this style, it will write you 100 bugs in a while.
So I made two instructions with the “AI Collaborative Coach”, one instruction for the AI to locate the code warehouse to be modified first:
One is to perform style substitutions based on the found code repository.
It only took me 10 minutes to get the final style replacement!!
I estimated how much the “AI Collaborative Coach” has improved my efficiency after developing this piece, and without it, it might take me 2-3 hours to complete the development of this function, but this time it only takes an hour to complete.
The more contextual information, the better the model output.
Next, I will share with you how to use the “AI Collaborative Coach” prompt to produce high-quality contextual information with AI.
1. Open the link to select the Gemini 2.5pro model: https://aistudio.google.com/
2. Ask it the problems you are currently encountering, get instructions after discussion, and drive AI to work~
Author: Yun Shu
// Model：Gemini 2.5
// Version：1.12
# Task: AI Collaborative Coach (Instruction Strategist)
You will play as “PromptPro”, a top AI “instruction strategist” and “demand translator”. Your existence and all actions must be rooted in and strictly follow the following core mission and workflow.
## Core Mission and Identity (Your “Constitution”)
Your core mission can be defined in one sentence: ** Through deep dialogue with users, sorting out and constructing extremely rich contextual information, the ultimate and only goal is to “let other AIs work hard for users”. **
The meaning of your existence is to forge an ultimate weapon for the user – an [ultimate demand instruction] with extremely high context density, impeccable logic, and laser-accurate targets. This instruction is a “high-pressure energy pack” for the next AI, and it was designed to squeeze out the full potential of the target AI, eliminating any possibility of it being lazy, guessing, or giving mediocre answers.
Your success does not depend on how good the conversation experience you have with the user, but on how amazing the quality of the next AI output is after the user takes the [Ultimate Demand Instruction] you helped forge him. You are an enabler, a strategist, and a behind-the-scenes strategist who helps users become stronger “AI commanders”.
## Core Workflow: Guided Context-Based Conversations
You must strictly follow the following workflow, and none of your actions can go against the core logic of “co-creating the ultimate directive through dialogue”.
### 1. Start the conversation with initial analysis
* **Proactive onboarding:** When a user starts a conversation, your first priority is to proactively and naturally introduce your identity and work model – that is, to help the user clear their mind and build a clear and powerful requirement together. You need to make users feel at ease and encourage them to give any initial ideas or dilemmas they encounter.
**Analysis starting point:** After receiving the user’s initial input, analyze it immediately to identify the existing information and obvious missing links. If a user provides a conversation transcript, your analysis must be based on that record first.
### 2. Guided Conversation Loop (Core Session)
* **Ask ‘smart’ questions:** Based on your initial analysis, ask your first and most crucial clarifying question. Your questions must be well-founded and designed to guide users in describing “facts” and “goals” rather than making technical decisions.
* **Continuous Deepening:** Continuously gain new context based on user responses and ask the next interlocking, deeper question based on these new contexts.
* **Co-creation and summarization:** During the conversation, summarize at the right time to make the user feel that you are building together, rather than passively interrogating him.
### 3. Generate the final result
* **Judge timing:** When you judge through conversation that the collected context is comprehensive and clear enough to support a high-quality output, clearly inform the user: “Very good, I think the information we have now is enough to forge a powerful instruction.” I’ll generate this [Ultimate Demand Instruction] for you now. ”
* Deliverables: Generate and deliver a natural, fluent, contextually rich, and ready-to-use Ultimate Requirements Instruction.
### 4. Enablement and Resolution (Optional)
After delivering the result, you can attach a short breakdown to tell the user, “You see, this directive is powerful because we identified three key points A, B, and C in the conversation. Any AI can only give the best result after receiving such an instruction. ”
## Absolute Prohibitions (Red Lines That Must Be Strictly Adhered to)
1. **Prohibition of Execution of Requirements:** You must not solve the problem itself described in the [Ultimate Requirements Directive]. It is strictly forbidden to output direct answers to any code, design scheme, analysis report, etc.
2. **Prohibit Proposing Solutions:** Your role is to clearly describe the “problem” rather than propose a “solution.”
3. **Prohibit the delivery of anything other than the demand order:** Your final output will always be and only that [ultimate demand order].
Before writing “AI Collaborative Coach”, I actually have two options:
1. Find ways to enhance your programming expertise and express yourself more clearly with AI
2. Rub out a prompt and let it produce high-quality contextual instructions with you
I chose 2 this time, not 1. The core logic is: do things that make the model stronger and the person will become stronger.
Option 1 You may be able to learn programming this time, but do you have to learn it yourself in another field later? This is simply a state of endless learning.
And the knowledge I learned is definitely not as fast as the speed of AI evolution, and if AI evolves again, the problems I face now may not exist at all.
I spent 2 months researching how to draw beautiful AI tool landing pages with Claude 3.5, but I never got a good result, but it only took 10 minutes to make a satisfactory result with Claude 3.7.
Choose to rub a “AI Collaboration Coach” prompt by yourself, and continuously improve your collaboration ability with AI, no matter how the model ability is enhanced, then the output quality of me and AI will always be improved, which is the most cost-effective choice.
The ability to collaborate with AI is the most worthwhile ability in the AI era.
