Applying the V-model to Genspark AI Development - The Debut of the Eric-George Method

Introduction

"The same kind of bug appeared again... Even if I fix it, another bug occurs somewhere else."

During the development of a web application using Genspark, I was facing a serious problem. The pre-rendering implementation for SEO was mistakenly applied to the homepage, causing the fortune-telling function to completely stop. Although the fix itself was completed in about a day, I was fed up with the AI introducing bugs without considering the scope of impact.

Just when I thought I had finally restored the fortune-telling function, two new bugs were discovered. To break this cycle, I decided to apply the traditional software development methodology called the V-model to AI development.

Background of the Incident: A Chain of Recurring Bugs

Misapplication of Pre-rendering Implementation

The root of the problem was the SEO measure implemented in v2.17.3 on December 15, 2025. At that time, pre-rendering (server-side rendering) was introduced to index blog posts in Google Search.

However, this implementation was mistakenly applied to the homepage (/). The homepage has a fortune-telling function that runs as a React SPA, and pre-rendering prevented JavaScript from executing, causing the fortune-telling function to stop completely.

The Chain of Recurring Bugs

The cause of the vicious cycle, where new bugs emerged with each fix, was clear:

• Lack of specification understanding: AI overlooked crucial parts of the specification
• Misunderstanding of impact scope: Did not consider the impact of fixes on other functions
• Insufficient testing: Inadequate post-fix verification

The Idea of Introducing the V-model

What is the V-model?

The V-model is a traditional quality assurance method in software development. It clearly separates the development process into upstream phases (requirements definition, design) and downstream phases (implementation, testing), ensuring quality by establishing corresponding testing stages for each phase.

The advantage of this model lies in the philosophy of building quality into the design phase. Bug fixes after implementation require significant effort, so thorough analysis and design in the upstream phases are crucial. For more details, please refer to Fundamentals of Software Quality Assurance.

Applying to AI Development: The Eric-George Method

I considered applying this V-model to Genspark's AI development:

• Upstream Phase AI (named Eric for now): Thorough reading of specifications, requirements analysis, root cause identification, creation of implementation instructions
• Downstream Phase AI (named George for now): Faithful implementation according to instructions, testing, deployment

Ideally, they would operate as separate AI agents, but Genspark did not allow one AI agent to launch another AI agent.

This persona switching method is explained in detail in a separate article: Understanding the Decisive Difference Between Code Generation and Code Execution in Genspark.

Implementing the Eric-George Method

Afternoon, January 8, 2026: The Debut Begins

Although the fortune-telling function was restored in v2.18, I applied the Eric-George method for the first time to address two newly discovered bugs.

Discovered Bugs

• Bug 1: Blog post style is broken, Markdown symbols (##) are displayed as-is
• Bug 2: Blog posts have not been updated since January 4

Detailed Analysis of Bug 1: Markdown Conversion Issue

Step 1: Eric Mode (Upstream Phase)

First, I thoroughly analyzed Bug 1 in Eric Mode.

Observation of Symptoms

Code Analysis

Upon checking the renderBlogPost() function in webapp/functions/_middleware.ts, I found the problematic code:

// [Problematic Code] v2.18
function renderBlogPost(post) {
  // ... Omitted ...
  <div class="content">${post.content}</div>
  // ← Markdown is output as-is!
  // ... Omitted ...
}

Database Confirmation

Next, I checked the actual data in the blog_posts.content column of the D1 database:

## Sensitive A-type and Free B-type Individuals: A Combination That Stimulates Each Other and Opens Up New Worlds

The sensitivity of A-type individuals and the free-spirited nature of B-type individuals might seem incompatible at first glance.

However, it is precisely this combination that often serves as an opportunity for them to open up new worlds for each other.

Step 2: Creating Implementation Instructions

In Eric Mode, I created clear implementation instructions for George. These instructions would significantly influence the subsequent success.

Step 3: George Mode (Downstream Phase)

Switching to George Mode, I implemented according to the instructions.

Implementation Code

/**
 * Simple Markdown → HTML conversion function
 * @param {string} markdown - Content in Markdown format
 * @returns {string} Content in HTML format
 */

Returning to Eric Mode, I rigorously verified the implementation results.

Results: Partial Success and Revealed Challenges

Points of Success

Metric	Result
Bug 1 fix time	✅ Approx. 1 hour (previously several days to weeks)
Fix cycle	✅ Completed in 1 round trip (previously multiple times)
Impact on homepage	✅ Zero (fortune-telling function operates normally)
Root cause identification	✅ 100% (lack of Markdown conversion)
Documentation creation	✅ Implementation instructions created

The Reality of Bug 2: User Intervention Was Needed

However, for Bug 2 (blog post update stoppage), a serious problem was revealed.

Eric's Analysis and Oversight

In the initial analysis in Eric Mode, the following investigations were conducted:

1. Checked blog post list API → Latest post was January 4
2. Checked Cron Worker code → No issues with the code itself
3. Manually executed API → 1st attempt 503 error, 2nd attempt successful

Eric concluded that it was a "temporary outage of the Gemini API used for article creation" and proposed establishing a monitoring system.

Only after I personally checked the specifications, reviewed the Cron logs, and accessed the Cloudflare dashboard using the authentication information, did the true cause become clear.

Clarification of Abilities: Mid-level Employee vs. Middle Schooler Level

From this experience, the capabilities of Eric and George became clear.

Related article: Limitations of Judgment Revealed in Genspark AI Development

Lessons Learned: AI's Limitations and Next Steps

Effectiveness of the V-model

The Eric-George method was not a complete success, but its effectiveness was proven. By separating the upstream and downstream phases:

• ✅ Improved accuracy in root cause identification (accurately discovered the Markdown conversion issue)
• ✅ Stable implementation quality (no issues with George's implementation)
• ✅ Clarified understanding of impact scope (zero impact on homepage)
• ✅ Documentation is naturally generated (implementation instructions remain)

Limitations of AI's Judgment

However, at the same time, it also became clear that AI's judgment has serious limitations.

Necessity of Quality Check: The Genesis of Gemini API Discovery

Realizing that Eric's judgment was unreliable, I concluded that a third-party quality check mechanism was necessary.

Initially, I tried Genspark's Deep Research and text generation AI, but neither was suitable for quality checks. Therefore, I myself asked Eric, "Can the Gemini API used for article creation also be used for quality checks?", and found that it could.

Subsequently, under my initiative, I developed a quality check method called "Gemini QA Framework" using the Gemini API.

Details of the Gemini QA Framework will be explained in the next article: Gemini QA Framework Built with Genspark - Automating AI Quality Assurance

Related article: Demonstrating the Necessity of AI Quality Assurance with Genspark

Detailed Verification in the Next Articles

In this article, I recorded the "debut experience" of the Eric-George method. However, from the perspective of a quality assurance engineer, a deeper analysis is required.

Future Detailed Verification Themes

In the next articles (articles 51, 46), the following themes will be verified in detail:

Investigating the True Cause of Bug 2

In this article, I recorded the initial analysis of Bug 2 as "a temporary outage of the Gemini API" and the fact that "user intervention was needed." However, I have not yet answered the question of why the AI could not identify the root cause.

Article 52 will reveal the path leading to the development of the Gemini QA Framework through these verifications.

Conclusion

The debut of the "Eric-George method," applying the V-model to Genspark, was a partial success.

✅ Points of Success

• Fixed Bug 1 in about 1 hour (similar issues previously took several days to weeks)
• Proven the effectiveness of separating upstream and downstream phases
• Confirmed George's (implementation) capability to be at a mid-level employee level
• Documentation (implementation instructions) is naturally generated

❌ Remaining Challenges

• Eric's (judgment) capability is at a middle schooler level
• Bug 2 required user (my) intervention
• AI does not proactively suggest available tools (Gemini API)
• Even when claiming to have "read the specifications," crucial items are actually overlooked
• Omits basic verification procedures (log review, authentication information reference)

This experience will evolve into the subsequent development of quality assurance methods, phase-based checks, and the establishment of approval workflows. It was a valuable debut, where I realized the importance of learning from both failures and successes and continuously improving.

---

If this article was helpful, please take a look at our other articles. We will continue to share practical insights into AI development using Genspark.