Vibe Coding: Redefining Product Management in the Age of AI

In this era where AI is reshaping productivity, the DIA project has disrupted traditional software development logic in just one month. This article deeply analyzes how to reconstruct scheduling and expense management systems using AI-native thinking, revealing key technological breakthroughs like Agent architecture and planning pipelines, providing product managers with a comprehensive methodology for AI transformation from cognition to implementation.

In this era, effort is the least valuable variable.

If you are still submitting tickets to developers for changing a line of text in an app, or if you find yourself staring at countless screenshots of competitors at 1 AM, mechanically filling out forms in Excel; if you think that “understanding technology” means writing every line of lengthy business logic yourself—then you are being downgraded by this era.

A month ago, I embarked on an experiment. As a non-technical builder, I attempted to challenge a field considered a “relic of the old era”: scheduling and expense management. In a month, I wrote over 100,000 lines of raw code to replicate an extremely complex, cross-domain AI-native product (DIA).

This was not a traditional development process, but a deep dive into “Vibe Coding.” When engineers at Anthropic, aided by AI, achieve productivity rates 1000 times that of Google engineers, we must realize: the old walls of software engineering have crumbled. Today’s developers are no longer “code farmers” but “commanders”; code is no longer a chain that binds creativity but a “Vibe” that flows freely.

1. Cognitive Awakening: Why Are Web 2.0 Products Becoming “High-Level Paralysis”?

Before diving into practical operations, we must break through a layer of glass: 90% of the tools on your phone are essentially “high-level paralysis”.

1.1 The End of Form-Based Interaction

Think back to how you create a calendar or bookkeeping entry: click the plus sign -> fill in the title -> select time -> choose category -> click confirm.

“By the time you finish filling out this form, you may have forgotten what you were trying to do.”

This structured input based on “forms, icons, and graphical interfaces” is due to the old databases being able to recognize only rigid data. The emergence of large models (LLM) allows us for the first time to achieve “fuzzy expression, precise execution”.

1.2 What Is a True “Native AI Product”?

Many products add a chatbot and claim to be AI products. But that is merely an “old-era product wrapped in an AI shell.” They are isolated from the underlying database and cannot achieve seamless operation. A true AI-native product (DIA) should be:

Agent = Large Model (Brain) + Memory (Memory) + Planning (Reasoning) + Skills (Hand/API).

LUI (Language Interaction) comes first, GUI (Graphical Interface) comes later.

2. Deep Deconstruction: The Underlying Architectural Logic of Vibe Coding

In this month of practical experience, I developed an efficient architecture suitable for Vibe Coding. This is not just a technical choice but a reconstruction of product thinking.

2.1 Thin Client Principle

The primary principle I followed is: The iOS side only does “dumb rendering.” All CRUD (Create, Read, Update, Delete), Agent loops, and tool calls occur on the backend. The iOS client acts like a waiter; it doesn’t even know what it is serving; it only renders the interface according to backend instructions.

Value: This greatly reduces the cost of multi-end adaptation. If you want to expand to Android or PC, you only need to write the front-end interaction, while the core business logic (cloud) remains unchanged.

2.2 Domain-Driven Design (DDD): High Cohesion, Low Coupling

In DIA, I divided “scheduling management” and “expense management” into two independent business domains.

• Mutually Non-Polluting: Managing time requires location and people; managing money requires amount and category. They operate on their own logical tracks.
• Cross-Domain Collaboration: The top-level “orchestrator Agent” issues tasks simultaneously. For example: “Help me reimburse the invoice for the business trip,” the orchestrator will simultaneously activate the scheduling Agent (to find the business trip event) and the expense Agent (to handle receipt recognition).

2.3 “Planning and Execution Pipeline”: The Ultimate Efficiency Flywheel

This was the most significant architectural adjustment I experienced during development.

• Initial Phase (LLM Parser): Only let the model parse intentions, relying on handwritten scripts for execution. The drawback is the loss of flexibility, unable to handle cancellations or modifications midway.
• Growing Pains (ReAct Logic): An “inference-execute” infinite loop. The downside is that batch operations are extremely slow; deleting four schedules requires going through four loops.
• Mature Phase (Planning and Execution Pipeline): The model first generates a complete “plan” containing N steps, presents it to the user for confirmation, and the database executes it in milliseconds all at once.

This reduced the time for batch operations from 2 minutes to just a few seconds.

3. Practical Pitfalls: How to Navigate “Imperfect” AI Programming?

Vibe Coding does not mean you can completely relax; it demands higher requirements for “observability” and “prompt engineering”.

3.1 Solving the “Mystical Bug” of AI Code

AI writing code (especially native iOS code) often results in hard-to-trace “mystical errors”.

Solution: Use Web Languages (Bridge Solution). In complex interactions like calendar views, I chose to bridge hardware capabilities with web pages. Because the training data for web languages is more extensive, the quality of AI-generated code is significantly higher than that of Swift/SwiftUI, greatly enhancing development efficiency.

3.2 Establishing an “Observability” Tracking System

In over 100,000 lines of code, nearly 30% is “monitoring code” unrelated to functionality.

Logging ID Tracking: Assign a number to each behavior of the Agent and each transmission link.

Pinpointing Bugs: When the system is unresponsive, you don’t need to read the code yourself; instead, let the model scan specific log numbers. This is 8-10 times faster than having AI blindly scan all code.

3.3 Self-Improvement Flywheel: Errors as Test Cases

I established an automated closed loop: every error in the production environment automatically converts into a test case in the local environment.

“Before every deployment, I spend an hour running nearly a hundred automated cases, watching the simulator automatically generate cards; that sense of ‘certainty’ is the cornerstone of Vibe Coding’s success.”

4. Extreme Aesthetics: Tokenized Design System

Even an AI-assisted development tool can fail if it looks “ugly.”

I did not adopt the clumsy method of “drawing a picture to implement a page”; instead, I established a system of over 10,000 lines of “Design Tokens.”

• Symbolic Definition: Abstracting fonts, spacing, materials, and frosted glass effects into Tokens.
• Rapid Replication: Once the specifications are set, adding any new component only requires telling AI to “read the current specifications” to generate visually consistent UI.
• Layout Logic: In today’s mobile interaction where patterns are largely established, do not try to challenge user habits. Layouts follow industry benchmarks (like Timepage) while innovating in color schemes and branding.

5. Methodology: How to Start Your First Line of “Vibe Code”?

As a product manager, you don’t need to become an algorithm scientist, but you need to be an excellent “architect builder.”

1. Identify Scenarios That Can Be AI-ified

   • Look for inefficient processes that require a lot of “filling forms, clicking, searching, organizing.”
   • Assess whether there are “cross-domain” needs (e.g., scheduling + expenses, scheduling + emails).

2. Embrace SDKs and Open Source Frameworks

   • Don’t reinvent the wheel from scratch.
   • Agent Development: Recommended frameworks like LangChain or Microsoft Semantic Kernel (noted as MSDK) can help you manage tool loops, context management, and structured outputs.
   • Database: Even if you are just doing a demo initially, consider scalability (multi-tenancy, security) in your database selection to avoid future technical debt.

3. Enhance Instruction Precision

   • Smart Agent, Dumb Tools: This is a core principle. Keep tools (APIs) extremely simple and pure (only for execution), while leaving validation, parsing, and reasoning to the Agent with global context.

4. From “Dialogue” to “Closed Loop”

   • First, use AI to run through the logic demo.
   • Introduce observability code.
   • Build an automated regression testing system.

Conclusion: Reclaiming Creativity

The title of Software Engineer is becoming blurred, replaced by “Builder.”

In a highly developed internet environment, there are no absolute demand vacuums. What we need to think about is how to provide a solution that is ten times better than the old era using new era (Agent), new tools (Vibe Coding), and new thinking (LUI).

As Karpathy (a founding member of OpenAI) said, the vast field of code generation has largely been solved. Next, it is the golden age for “non-technical background” builders who have drive, ideas, and are willing to spend time mastering AI tools to change the world.

If you can, then start now.