The Core Technique of Vibe Coding: Give Your Judgment Criteria to AI

The Biggest Mistake Most People Make with AI Programming

The most common error people make when using Vibe Coding is telling AI what to do without explaining the reasoning behind it.

Vibe Coding is a concept proposed in 2024 by Andrej Karpathy (OpenAI co-founder, former Tesla AI Director). It refers to developers completing coding work through natural language conversations with AI programming assistants (such as Cursor, GitHub Copilot, Claude, etc.) rather than writing code line by line. In this paradigm, the developer's role shifts from "the person who writes code" to "the person who guides AI to write code," with the core competency moving from syntax proficiency to precision in requirement expression and architectural judgment.

For example, a seemingly normal instruction like: "Help me optimize the homepage" — this sentence hides all the critical information. Who is the homepage for? What's the biggest problem right now? Is this a minor tweak or a major overhaul? What style can't be lost? Which parts shouldn't be touched? What level of completion counts as good enough?

All of this exists in your head, but AI doesn't necessarily know it.

The quality of collaboration with AI doesn't depend on how long your Prompt is, but on whether you've handed over your judgment criteria. This touches on a common misconception about Prompt Engineering: many people believe the key lies in using specific magic words or fixed templates, but research shows that the output quality of large language models depends more on the completeness of contextual information — that is, how many constraints you provide to help the model narrow down the solution space. Judgment criteria are essentially the natural language expression of these constraints, and they're the key clues that upgrade AI from "generic execution" to "precise collaboration."

Six Judgment Dimensions You Must Include When Assigning Tasks to AI

These six dimensions essentially correspond to classic frameworks in software engineering requirements analysis: the goal corresponds to user stories, background to business context, pain points to problem definition, scope to iteration range, style to non-functional requirements, and boundaries to technical constraints. In traditional development, this information is conveyed through PRD documents, standups, code reviews, and other processes. In AI programming, it needs to be efficiently compressed and communicated within a single conversation.

First: What exactly needs to change

Don't just tell AI to "optimize the homepage" — clarify the expectation. For example: "Help first-time visitors understand more quickly that this is an AI collaboration tool."

Once this is stated, AI's attention shifts from "make it look better" to "make it clearer." Different goals lead to completely different outputs.

Second: Why it needs to change now

For example: "Conversion rates have been low recently — users see the first screen and don't know where to click next." This context is extremely important. Without context, AI will easily give you a generic optimization plan; with context, it knows this is about solving a conversion problem, not continuing to adjust visual design.

Third: What's the most frustrating issue right now

For example: "The headline is too abstract, and the three feature points all read like marketing copy." This is far more useful than "it doesn't feel premium enough."

"Doesn't feel premium enough" leads AI into an aesthetic dead end, while "abstract headline, feature points sound like marketing copy" are things that can actually be fixed. Give AI specific pain points, and it can give you specific solutions. The principle behind this: large language models need clear, actionable constraints to converge their output space. Vague feeling-based descriptions cause the model to wander randomly across multiple possible directions, while specific problem descriptions point directly to solution paths.

Fourth: The scope of this change

For example: "Just a small tweak — don't redo the page structure." This information is crucial. AI can easily turn a 30-minute task into a half-day refactoring project — not because it's lazy, but precisely because it's too eager to solve problems. If you don't define the scope, it will tend toward giving the most thorough solution possible.

This phenomenon has technical reasons: during training, large language models are exposed to vast amounts of best practices and refactoring examples. When they identify optimization opportunities in code, they tend to provide the most complete solution. In software engineering, this is called "scope creep," and it's especially pronounced in AI programming scenarios because AI has no concept of time cost and won't automatically weigh the balance between "benefit of change" and "risk of change." Defining scope helps AI establish this trade-off framework.

Fifth: The flavor that can't be lost from the project

For example: "Keep the current restrained, tool-like style — don't turn it into a SaaS landing page."

This is ten times more useful than "make it more premium." Because "premium" is too broad — some people think whitespace is premium, others think black and gold is premium. If you don't specify, AI can only draw from a grab bag of generic aesthetics. Clarifying style preferences keeps the output stable in the direction you want.

Sixth: What areas to leave alone

For example: "Don't touch the login flow, don't swap the UI library, don't change the API." You need to tell AI what its boundaries are. An AI without boundary constraints is like a construction crew without fences — they might tear down walls that shouldn't be torn down.

Complete Example Prompt: All Six Dimensions in One Shot

Combining all six pieces of information above, a practical Vibe Coding instruction looks something like this:

Help me revise the homepage hero section. The goal is to help first-time visitors understand more quickly that this is an AI writing tool. The current problem is that the headline is a bit abstract, and the feature points lean too much toward marketing copy. Just a small tweak — don't redo the page structure. Keep the current restrained, tool-like style — don't turn it into a flashy SaaS landing page. Start by looking at the Home and Hero related files, and don't introduce any new UI libraries.

This paragraph already contains all six dimensions:

Dimension	Content
What to change	Help users understand the product positioning faster
Why change it	Headline is abstract, feature points sound like marketing copy
What's frustrating	First screen doesn't communicate clearly
How much to change	Small tweak, don't redo structure
What to preserve	Restrained, tool-like style
What not to touch	Don't introduce new UI libraries

It's worth noting that while modern AI programming assistants have large context windows (ranging from tens of thousands to hundreds of thousands of tokens), a large context window doesn't equal high information utilization efficiency. Research shows that models pay more attention to information at the beginning and end of the context (the "Lost in the Middle" phenomenon). Therefore, organizing the most critical judgment information — goals, constraints, boundaries — in a concise and powerful way, as shown above, is more effective than writing a lengthy but loosely structured Prompt.

Core Insight: AI Execution Quality Depends on the Judgment Clues You Provide

The comfortable state of collaborating with AI isn't about writing a ten-thousand-word template — it's about leaking a bit of the judgment in your head ahead of time.

The quality of AI's execution and completion depends entirely on the judgment clues you give it. When you provide a bit of background, scope, preferences, and boundaries, it can truly "join" the project rather than acting as an outsider giving generic advice.

This approach applies not just to frontend page optimization — it works for virtually all AI programming scenarios. Whether you're writing backend logic, refactoring code, or designing database structures, you need to hand over the judgment of "why do it this way," not just the action of "what to do."

Remember: Anyone can state actions, but judgment is where your value lies — and it's the input AI needs most.

Key Takeaways

• The most common mistake in Vibe Coding is only telling AI the action (what to do) without providing judgment criteria (why to do it this way)
• When assigning tasks to AI, supplement with six key pieces of information: what to change, why to change it, what's frustrating, scope of change, what flavor to preserve, and what to leave alone
• The difference in AI collaboration quality isn't about Prompt length — it's about whether you've handed over your judgment criteria
• Specific pain point descriptions (like 'the headline is too abstract') are far more effective than vague feelings (like 'it doesn't feel premium')
• Defining clear boundaries and scope prevents AI from over-refactoring, turning a 30-minute task into a half-day project