Three Months as an AI Code Babysitter: My Exhausting Journey and Hard Lessons

1. The "Easy Money" Project

In February 2026 a client from the offline retail chain asked for a "store inventory + membership points + mini‑program" system for 50,000 CNY and a 1.5‑month deadline. The author, a 14‑year Java veteran, believed AI could turn this into a hands‑free profit.

2. First Month – AI Generates Code Faster Than a Human

The AI (unnamed) accepted plain‑text requirements and instantly produced a compressed project archive. Examples:

"Generate a Spring Boot + MyBatis Plus inventory backend" → project downloaded in 3 minutes.

"Add a membership points module, 1 CNY = 1 point" → seven or eight classes were added.

"Generate a Vue 3 front‑end" → front‑end code appeared with a single click.

During a demo the client exclaimed, "Your progress is too fast—are you working overtime?" The author posted proudly on social media, claiming AI made him ten times more efficient.

3. Second Month – New Requirements Break the Generated Code

On 2026‑03‑15 the client added four major features: inter‑store transfer, WeChat‑integrated points, a sales‑analysis dashboard, and fine‑grained role permissions. The author tried to extend the existing AI‑generated code but encountered three major nightmares:

3.1 Project‑Structure Inconsistency

The AI used three different service patterns across modules: XxxService +

XxxServiceImpl

XxxManager

(no interface) XxxBiz (business logic class)

These styles coexisted in a single codebase, making maintenance impossible.

3.2 Chaotic Database Schema

Fields for the same concept used three different names, types, and value sets (e.g., member_status int, order_state varchar, is_active tinyint). None of the 52 tables had foreign‑key constraints; all relationships were hard‑coded with if‑else. Column names mixed Chinese pinyin and English abbreviations ( huiyuan_jf, mddh).

3.3 Security Vulnerabilities

SQL built by string concatenation, exposing injection risks.

Plain‑text password stored as varchar(50) with value 123456.

File‑upload endpoint accepted .jsp without validation.

JWT secret hard‑coded as "123456".

The author warned that deploying such code would land the client in the news within three days.

4. Third Month – The Code‑Cleaning Phase

From March late to April end the author spent every day refactoring:

Morning: reviewing AI‑generated “half‑finished” code.

Mid‑morning: redesigning all 52 tables, adding naming conventions, foreign keys, indexes, and comments.

Afternoon: unifying layers, exceptions, logging, and return formats; removing all XxxManager and XxxBiz classes.

Evening: fixing vulnerabilities – converting string SQL to prepared statements, encrypting passwords with MD5, adding file‑type whitelist, and tightening permission checks.

Late night: rewriting every unit test from generic assertNotNull(result) to meaningful assertions.

The workload caused a 4‑jin weight loss and doubled hair‑loss, prompting the author’s wife to ask whether the project cost 5 wan or 50 wan.

5. Turning Point – Discovering FeiSuan JavaAI

After venting in a group chat, a colleague shared the FeiSuan JavaAI product page (https://www.feisuanyz.com/home). The marketing slogan claimed “from process‑driven to autonomous collaboration, multiple expert agents work transparently.” Skeptical but desperate, the author installed the IDEA plugin.

6. Five‑Step Multi‑Agent Workflow

FeiSuan JavaAI splits development into five agents:

Step 1 – Requirement Planning Agent

It parses the “store transfer” requirement, discovers existing tables ( t_store, t_product, t_inventory), and outlines six APIs (initiate transfer, approve, inventory flow, shipping, receiving, transfer record).

Step 2 – Interface Design Agent

Generates detailed API specifications for each of the six endpoints.

Step 3 – Database Architecture Agent

Reads the connected MySQL schema (databases prefixed with db_, tables with t_) and creates only two new tables: t_store_transfer_order and t_store_transfer_item, establishing proper foreign‑key relationships and using consistent naming ( datetime, tinyint with enum comments).

Step 4 – Business Logic Agent

Lists every business rule, approval flow, and inventory state transition, ensuring the generated logic matches the original requirement.

Step 5 – Source Code Generation Agent

Produces clean, convention‑compliant Java code (Spring Boot + MyBatis Plus). Passwords are stored with MD5, SQL uses parameterised queries, timestamps follow yyyy‑MM‑dd HH:mm:ss, and all generated code matches the project’s existing style.

Crucially, every action is logged, so the author knows why each decision was made.

7. The Ten Expert Agents

Beyond the five core agents, FeiSuan provides ten specialist agents that rescued the “AI‑generated mess”:

One‑Click Fixer : Detects 23 compile errors, 41 null‑pointer risks, 7 unclosed streams, and 13 deprecated API usages; fixes them automatically.

Security Fixer : Reports 19 SQL‑injection points, 6 XSS, 3 plain‑text passwords, 11 hard‑coded secrets, 2 unchecked uploads, 5 missing CSRF protections; applies one‑click patches (e.g., replaces string SQL with prepared statements, adds MD5 hashing).

Unit Test Generator : Generates meaningful tests covering normal, exception, and edge cases; raises coverage of the transfer module from 0 % to 87 %.

Project Documentation Generator : Emits a 30‑page Word document with architecture, module descriptions, API list, DB design, and deployment steps.

Java Cleaner : Reformats code to Alibaba/Google style, splits deep nesting, and optimises imports.

Dependency Fixer : Detects a NoSuchMethodError and suggests upgrading Spring Boot 3.x with MyBatis Plus 3.5.5+, generating the migration plan.

Framework Upgrader / Migrator : Generates a full migration from Spring Boot 2.7 to 3.2 (jakarta namespace, annotation updates, config migration).

Best‑Practice Optimiser : Analyses code against official framework recommendations and offers one‑click improvements.

8. SQL Chat – The First Truly Business‑Aware AI

The original AI produced a 27‑level nested query with obvious injection points. SQL Chat, bound to the local MySQL schema, answered a natural‑language request:

"Find the top‑3 selling products per store and category since May 2026, ordered by sales amount."

It returned a concise query using

ROW_NUMBER() OVER (PARTITION BY store_id, category_id ORDER BY sales_amount DESC)

and suggested creating a composite index on t_order_detail(store_id, category_id, sales_amount). It also confirmed the query was parameterised and injection‑free.

9. Final Outcome and Advice

The original 50,000 CNY project was delivered two weeks late, the client paid the final invoice and referred a second project. Using FeiSuan JavaAI, the author rebuilt the entire codebase, reduced the estimated effort for the new project to 60 % of the original forecast, and saved countless hours of manual debugging.

He concludes with a warning: "If you are still being tormented by end‑to‑end one‑click AI code generators, let AI become your team instead of you becoming its babysitter."