AI Productivity – AI Coding + Product Development: Faster, Different, Better? (with Gemini CLI)

The era of AI coding has arrived. Currently, there are numerous AI coding tools available on the market. They began to emerge around a year and a half ago, and the past year has seen development rapidly. From traditional IDEs to IDE extensions and plugins, and now even abandoning IDEs entirely to use prompt-based coding directly in browsers, a variety of approaches coexist.

Table of Contents

閱讀繁體中文(台灣) 版本, 或可右上角選單中語言切換「中文(台灣)」

Feel free to get AI-generated Voice Summary in 2 mins via below podcast source.

–

For engineers, these developments represent significant changes and the establishment of new ways of working. But what kind of impact does this have on product managers and product development designers? Many believe that the emergence of new AI coding tools — particularly those that abandon the traditional IDE approach, such as Lovable,Replit and Base44 — could even allow product design to connect directly with programming. Could this really happen? Can product managers build the product’s codebase directly?

Previously, we tested using AI coding products like Lovable and Base44 for MVP development. This time, we completed a side project using Gemini CLI, which provided us with extra ideas and insights of this issue.

1. AI Coding Generation is Here. Tools.

Over the past year and a half, we have seen an massive growth in AI coding tools. Based on usage patterns and differences between tools, we can categorize them as follows:

Existing IDE AI Extensions、Plugin or Assistants — CodePilot, Google Code Assist, etc.
General-purpose AI Assistants (highly flexible integration with existing IDEs) — Claude Code, Gemini CLI, etc.
Next-generation AI IDEs — Cursor, Windsurf, etc.
Next-generation Vibe Coding Product(particularly browser-based solutions that abandon IDEs) — Lovable, Replit, Base44, etc.

Observations:

Most engineers have experience with category 1.
Category 2, when integrated with existing IDEs, can achieve effects similar to category 3.
Currently, the hottest categories are 2 and 4 (We have also tried category 4 tools (Lovable, Replit) earlier.)
Category 2 not only collaborates extensively in programming but can even go beyond coding, functioning as a comprehensive personal assistant (multi-agent, agent teams, etc.).
Category 4 eliminates the effort of using a traditional IDE and attempts to provide non-engineers with a more accessible way to create products.

2. Are Traditional Product Development Workflows Being Abandoned?

Many people are discussing how the next generation of product development processes will change. Here, we explore this question from the perspective of a product manager.

– What Are the Responsibilities and Workflows of a Product Manager?

Will the introduction of AI coding reshape how we approach product development and design workflows? What might the workflows of the future look like? Will engineers be replaced — or will it be product managers who be replaced?

First, we need to recognize that the responsibilities of product development and design go far beyond simply writing specifications for engineers or guiding the 0–1 development of products and features. Nor is every feature about creating visual, interactive functions. Product managers are still required to handle a broad range of tasks, including:

Design and conduct 0–1 MVP rapidly
Feature design focused on UI/UX (screen layout, interactions, and user experience)
Feature design beyond UI/UX (functional, security, or architectural needs such as API design, webhook design, component design, or system-to-system integrations)
Performance and architecture design (non-functional adjustments such as performance optimization, security handling, or structural improvements)
Legacy product development (refactoring, modernization, or ongoing maintenance)
Continuous iteration and updates for existing products or projects (not all products are built from scratch through AI coding)

The above responsibilities are still mostly related to coding part. However, a product manager’s role is not 100% tied to coding task planing. Product managers also handle a wide range of broader product-related tasks, including preparation, research, customer engagement, and operational responsibilities:

Brainstorming new ideas (new features, new product directions, new campaigns, etc.)
Roadmap planing
Visual design optimization and proposals (graphic, visual, motion, and branding improvements)
Competitor survey
Market research
Requirement and user interviews
Specification and requirement documentation (PRD, spec documentation)
User manual documentation
User feedback handling
Product performance monitoring (metrics monitoring)
User case studies
Stakeholder coordination and management
Risk identification and management
Operation management
Growth management
…

– The work of a product manager may evolve or be integrated, but it will not disappear.

The responsibilities of a product manager are not limited to the ones listed above, but even from this overview, we can already see how many different areas are involved. This also means that the entire product development and management covers a wide range of tasks.

Can all of these really be replaced by someone outside the PM role simply by using AI tools? Or will AI tools completely take over these responsibilities from these roles? Probably not.

Most of the work will not disappear — it will become more efficient or be integrated.

AI era, if someone replaces you, it will likely be they are using efficiency tools to perform your responsibilities.

3. What Are the Debates Around the New Product Development Process?

Let’s assume that the current debate around whether the introduction of next-generation AI tools in product development workflows will make the product manager role disapper (replaced by engineers). Conversely, the question could also be whether the engineer role will be replaced (replaced by product managers).

– Why Does This Debate Exist?

The reason this debate exists is that next-generation AI tools enable people who originally lacked coding skills to build software, in another side, also allow those without product design experience to directly generate visual interfaces from descriptions.

AI tools give you capabilities you didn’t have before, which makes people imagine you can take over more responsibilities.

AI tools have broad capabilities, which leads people to imagine that AI could completely replace existing roles.

Some people assume that, with next-generation AI productivity tools, the product development workflow might evolve in the following ways:

No product manager needed — engineers use AI tools to replace the product design task.
No engineer needed — product managers generate code directly into a final product using AI tools.

– What Would Happen in Practice After AI Is Introduced?

Based on the actual workflows we have recently tested, while these scenarios are possible, but somehow they will not replace roles 100%. In other words, the responsibilities of product managers and engineers may be expanded or integrated, but the work itself will not disappear.

Why is that? Looking back at the product development and design responsibilities we discussed earlier, since the work itself does not disappear, it will only be supplemented for efficiency by some roles.

If engineers step in, they are literaly performing the tasks of a product manager.

If someone tries to rely on AI to fill these roles, ask yourself: is there any task that AI can complete entirely without human involvement, oversight, or review? Since this is highly unlikely, it becomes clear that AI primarily enhances efficiency rather than replaces responsibilities.

| AI era, if someone replaces you, they are most likely using efficiency tools to perform your tasks.

| The work does not disappear, but that does not mean the number of positions will remain the same. Efficiency gains often lead to a reduction in headcount.

https://www.reddit.com/r/OneAI/comments/1l1c4xr/ai_it_job_postings_up_448_while_nonai_it_jobs/

4. Which Scenario Is the Most Realistic for Next-Version Product Development?

What might the future look like? Potential scenarios include:

Significant productivity gains for anyone leveraging AI tools.
The emergence of cross-functional capabilities — tasks traditionally tied to a specific role may be performed across roles.
The responsibilities of product managers and engineers will mostly remain, but their efficiency could be greatly enhanced.
New roles may emerge in some teams, such as a Product Engineer, but this will not lead to a complete replacement of engineers + product managers.

Significant Productivity Boost Across Roles
With the adoption of AI tools, the productivity of any role can be greatly enhanced.

What might the future look like? Potential scenarios include:

1. With the adoption of AI tools, the productivity of any role can be significantly enhanced.

Whether it’s product managers, engineers, or other team members, the evolution of the AI industry has already introduced tools that ultimately translate into improved efficiency.

2. With the adoption of AI tools, cross-functional capabilities are likely to emerge.

The generalization capability of AI and the application based on LLM, cross-functional skills will become more accessible. e.g. Engineers can prompt AI to design UI and get the Figma outputs via Figma MCP. Product managers can generate code projects directly through prompts.

3. The responsibilities of product managers and engineers would not disapper, but be significantly enhanced.

As discussed earlier, these duties are unlikely to disappear. The introduction of AI tools primarily serves as a means of boosting efficiency and extending capabilities. However, when it comes to orchestration, oversight, and fine-grained control, a human with contextual understanding is still essential to take the decision.
(We will illustrate this point later with an example project built using Gemini CLI.)

4. The Possibility of New Roles 「Product Engineer」, But Not Across The Board.

In some smaller teams, a new role — the Product Engineer (in the AI era) — may emerge. Essentially, this role consolidates the responsibilities of both product managers and engineers into one person. However, the underlying responsibilities and work requirements themselves do not disappear.
This means such a setup is more likely to be adopted by small teams or in specific domains, but it is unlikely to become universal. Consider a more extreme case: in theory, an AI agent team could replicate any role. Does that mean a single CEO with an AI agent team could operate a company that traditionally requires 100 people? Maybe, this will not happen.

5. Short-Term Contraction of Roles / Employee Headcount

While responsibilities are unlikely to disappear on a large scale, we are currently in the early phase of AI adoption — an era focused on efficiency gains. Over the next 1–2 years, the number of related roles may stagnate or even contract slightly, until organizations fully adapt to these efficiency tools.
The future of product development and design workflows will likely be a hybrid model, where teams integrate a variety of AI tools to drive efficiency, yet still perform most of the existing responsibilities. As teams experience the productivity benefits, companies may choose not to expand headcount — or may even reduce it — until efficiency becomes the new normal across the industry.

5. AI Coding+Product Development: Practical Case

We have already outlined the conclusions regarding the potential changes in product development and the role of product managers. Now, let us return to the practical utility and challenges of AI in coding and product development.

In this case, we used the Gemini CLI to collaborate on a side project. By going through the entire process — from initiation to development, and from development to actual operation — we identified both advantages and drawbacks.

*For another AI coding discussion (using Lovable), please check this post:
AI Coding an MVP in 1 Day (with Lovable)

https://medium.com/@neil.taiwan1999/ai-coding-an-mvp-in-1-day-with-lovable-5214c1f6a0fb?source=post_page—–d685b7279539—————————————

The following showcases the full process of using VS Code + Gemini CLI to complete a side project together with me, highlighting the situations encountered and the observations made along the way.

– Requirement: Build a personal focus journal system with support for various statistical methods.

I have a habit of keeping daily work notes. Normally, I use AnyType to edit and share a single record across both my phone and computer. I record, in plain text, the items I spend focused effort on each day. Sometime, I check them or doing quick calculations to understand whether my daily schedule needs adjustment or to remind myself to stay more focused.

Based on these text records in AnyType, I would like to build a suitable platform that can provide statistical analysis and better tracking.

– Specification Creation(1h)：Discuss with Gemini CLI to build it

We adopted a Text Spec Driven approach. After a few rounds of interaction and discussion with Gemini CLI, a file named “Requirement_Specification.md” was produced.

– Prototype Creation (0.5h): Using the “Requirement_Specification.md”, we initiated code generation.

I instructed Gemini CLI to build the corresponding project based on the specification, specifying React + NodeJS as the tech stack. After approximately 1–2 prompts, a reasonable prototype was generated. It met the functional requirements, had a sensible layout, and was executable. The prototype was very basic, but fully aligned with the specification.

– Layout、UI/UX、Debug、function Adjustments (10h)：

The initial prototype had several shortcomings in layout, UI/UX, and user interaction patterns. Throughout the process, we iteratively refined the project through multiple prompt adjustments. Gemini CLI was able to handle modifications based on semantic instructions in both Chinese and English for most cases.

Some issues arose from bugs generated by Gemini CLI itself, which needed debugging assistance (to be shared in detail later). Additionally, some adjustments involved structural changes unrelated to core functionality.

– Project Completion (Total: 12h)：Approximately 12 hours.

Finally I was able to use this small product’s focus journal recording and statistical features to conveniently review my own work time costs about AI Coding.

*My typical time-tracking format for this type of work is similar to:
“22:30 — Producing — Coding — Gemini — 00:30”

– Product Result：Demo (.gif)

6. AI Coding+Product Development: Real Issues

– Misconception 1: This represents the entirety of a product manager’s work

Although we were able to quickly build a product in a short amount of time, the product is very small. Examining the work involved, it actually only covers a portion of a product manager’s responsibilities.

Whenever we see others showcasing the impressive results of “Vibe Coding,” have we considered that this is often just the execution phase after requirements have been confirmed (or generating specifications along the way)? In reality, larger product companies have many more product management tasks that need to be carried out beyond this stage.

– Misconception 2: This represents the entirety of an engineer’s work

The same misconception applies when one thinks AI can fully replace engineers. While AI coding can accelerate the programming phase, engineers’ responsibilities also cover many foundational aspects that are not only visible part, including architecture, performance, security, and system integration.

AI coding indeed helps speed up part of the work, but it does not replace the full scope of an engineer’s responsibilities.

– 12 Issues in Reality:

– Issue 1: AI tools may fail, leading to chaos

Failures can occur due to token limitations, cost constraints, network issues, or the AI itself getting stuck in a loop. This can result in changes being left incomplete midway. In such cases, you might be able to prompt the AI to fix the issues, but sometimes the situation becomes messy and requires human intervention to clean up or assist.

– Issue 2: AI is not guaranteed; simple problems may fall outside its scope

In this case, a single extra character caused a syntax error. Although the mistake was easily visible to the human eye, it happened to fall outside the AI’s handling capabilities. Multiple attempts to prompt the AI for a fix were unsuccessful, and ultimately, the error had to be corrected manually by removing the character.

– Issue 3: AI may generate unexpected garbage

In this case, the AI produced continuously accumulating comment garbage. This issue is not limited to comments — meaningless or redundant code may also be generated. Periodic human inspection is sometimes required to identify and clean up such trash.

– Issue 4: AI can still misinterpret semantics, leading to unexpected results

This case demonstrates a clear example of ridiculous semantic misunderstanding. The intention was to swap the visual positions of certain components. While the AI did adjust positions, it modified the positions within the HTML file rather than the rendered layout positions as seen in the final output. Ultimately, human guidance was required to prompt the AI to correct this.

– Issue 5: AI can still misinterpret semantics, leading to unexpected results

Another example of ridiculous semantic misunderstanding occurred in this case. The goal was to add logs in batches and then display the latest date in the Daily Log. After multiple unsuccessful attempts, reviewing the code revealed that the AI interpreted “latest” as “last,” and then “last” as the last entry in the file. This misunderstanding reversed the intended effect of displaying the latest date. Ultimately, human prompts were required to correct the behavior.

– Issue 6: Potential calculation errors due to incorrect formulas

An incorrect calculation logic led to errors in precision, as demonstrated in this case. The input duration format was xxH yym. To compute the total daily time, the AI converted this format into a decimal format x.xH. For example, 01:20 would be converted to 1.33h, then occuring a precision issue. If using this method for summation results in minute values that are never whole numbers. For instance, a total of 8.59h is a loss of precision rather than a reflection of the original data. Ultimately, human prompts were required to correct the calculation logic.

– Issue 7: Inconsistent data processing by AI can cause problems

As mentioned earlier, the original duration data format was xxH yym. The minutes (yy) should be correctly interpreted as minutes, but the AI processed the same data inconsistently—some entries were correctly recognized, while others were misinterpreted, treating yy as hours. These inconsistencies ultimately required human intervention to correct the calculation logic.

Unexpected behaviors like this, where identical data is processed differently.

– Issue 8: Inconsistent naming rule by AI may create maintainability issues

Even for the same operation, such as saving a file, the AI sometimes chose the name “Save”, while in other cases it chose “Add”. Inconsistent overall naming rule occurred frequently, potentially introducing maintainability challenges in the codebase.

– Issue 9: AI may produce unstructural stuff, creating maintainability issues

As shown in the figure, the AI’s decisions regarding structural organization and naming lacked consistency or not well structure. Over time, this can accumulate into “dirty legacy” code, potentially introducing maintainability challenges for future development.

– Issue 10: AI sometime requires human assistance

For Gemini CLI, we granted access only to the current folder and files, but not to the OS、specific software or browsers. So, the AI often required human assistance for debugging, including performing tests and collecting logs.

– Issue 11: Work with AI, require version control

Currently, Gemini CLI does not have a built-in recovery function. Even if requested, it cannot revert changes (although this could be implemented in the future). Given that the AI may make unexpected changes, the operator must maintain a sensitive version control mindset, such as keeping backups to allow recovery when needed.

– Issue 12: AI may modify code or files outside the intended scope

In some cases, even for minor edits, the AI may unexpectedly change files beyond the intended target. For example, a modification intended only for File A could also affect File B, or altering Feature A may unintentionally alter Feature B also. When delegating modification scope to AI, it is important to recognize that it is not 100% accurate.

7. True Fact of AI Coding

– The Ever-Changing Snake Game

Go back to the classic snake game, many people trying Vibe Coding may begin with simple, well-known games and be impressed by the AI-generated quality and capabilities. But have you ever tried prompting the same command multiple times to achieve the same goal?

You will notice that, even though the results are beautiful, each run produces a different outcome.

AI is doing One Shot game base on probaility

Some might say this demonstrates AI’s diversity, but in reality, it’s a probability game. The idea of producing a complete result from a simple command is similar to Zero-Shot or One-Shot result: the AI is matching the most likely outcome while performing extensive guessing and filling in blanks.

In fact, every release demo from major large models also showcases a large number of Zero-Shot examples.

While this looks impressive, is this snake game really the style, gameplay, or “feel” you wanted? Not necessarily. The Zero-Shot(or One-Shot) result serves only as a quick draft — the real work lies in the iterative adjustments that follow to achieve the desired functionality, style, and interaction.

You could also consider “preparing very detailed specifications” to increase the accuracy of the Zero-Shot(or One-Shot) outcome. However, this shifts your actual working time to the specification-preparation phase, and somehow maybe it does not guarantee better results. In fact, sometimes a single well-crafted page of specifications can outperform a large set of overly detailed ones.

In reality, the process of AI coding still requires specifications work. These specifications may be prepared in advance, or gradually developed through a series of prompts during the process. Before building any product, clear goals and specifications are still necessary.

Therefore, the work of product managers and engineers has not disappeared, but AI does provide significant efficiency gains.

– What AI Coding Can Do:

A rapid self-productivity tool: Increasingly, people can create their own tools to assist themselves.
A fast prototyping tool: It allows you to quickly visualize and test ideas.
An accelerated coding tool: It speeds up routine edits and basic engineering tasks.
A peer-partner collaboration tool: It not only helps with coding but also accompanies you in thinking through problems.

– What AI Coding Cannot Do:

Can’t launch a completed product in just a few prompts: The great product still requires refinement and polishing.
Can’t Guarantee a completely error-free, bug-free, or hallucination-free outcome: Users must have the ability to identify and correct issues.
Can’t disapper the work of product managers or engineers: AI mainly serves to improve efficiency, not eliminate these roles.

– Key Considerations for AI Coding:

Human-in-the-loop awareness: Consider the need and degree of human oversight in AI workflows.
Basic programming knowledge is recommended: You need it to identify potential issues; otherwise, risks may go unnoticed.
Convenient for self-productivity applications, but full or saas products are still far away: Especially for foundational or underlying systems.
Maintain supporting .md files: Include requirement specifications and code structure documents to guide development.
Possible for irreversible issues: AI may make unexpected operations, modifications, or deletions if given full access to project files.
Coding and naming conventions may be inconsistent: Extra care is needed to enforce structure.
Use layered architecture and protect individual modules: Limit AI edits to specific areas.
Engage in discussion could be better than automatic changes, : Consider discussing changes with the AI and decide manually whether to implement them.

The result is quite interesting. We found for Product managers, empowered by AI Coding tools, are more likely to gain basic programming capabilities. However, the more refined and complete the outcome you expect, the more you will inevitably need the deeper knowledge and expertise that engineers possess.

8. What Is the Next?

Ultimately, the goal is to provide tools or processes that improve efficiency and make work more effective. Product managers still have many responsibilities, but there are numerous model-based tools that can assist in their work.

In the area of AI coding, I believe that at a minimum, every product manager can build their own efficiency tools and quickly experiment with technology or create prototypes.

Prototyping is not limited to AI coding. In the era of image and imagination generation, there are many possibilities to support this goal. For example, tools like Google Stitch, Veo3, and Nano can help demonstrate scenarios or build MVPs. Other tools can help quickly generate MRDs, BRDs, PRDs, wireframes, mockups, or Figma designs — covering daily product management needs.

Spec-driven AI coding is also worth exploring. Examples include AWS Kiro, as well as services and frameworks from Claude Code and Git, which offer corresponding support for specification-driven development.