Can AI Build a Complete Full‑Stack Website Without Writing a Single Line of Code?

Background

The AI programming field is booming, with many claims that AI will replace programmers. While AI can generate code snippets easily, fixing bugs in complex projects often consumes more time, leading to messy codebases. To test these claims, the author conducted an experiment: build a complete website using only AI, without writing any code.

Goals

The experiment had two objectives:

AI‑native design and development : Follow the emerging concept of AI‑native applications, which place large language models (LLMs) at the core of the system architecture, use natural‑language interaction protocols, multimodal perception, and agent‑based toolchain orchestration, and rely on a data flywheel for continuous model improvement (as described in Alibaba Cloud’s AI Native Application Architecture Whitepaper ).

Implement typical website functionality : Build a practical inventory‑management system that supports CRUD operations and data dashboards, ensuring the AI‑driven approach serves real business needs rather than being a mere tech demo.

AI‑Native Design in Practice

The author interprets AI‑native development as both a technical shift—designing the whole system around an LLM instead of treating AI as a peripheral tool—and a mindset shift, where interaction, design, and implementation all start from an AI‑centric perspective. The experiment reflects this by using AI for every step, from prototype generation to final code, and by making the AI chat assistant the primary user interface.

Prototype Design

The author started with Calicat, an AI‑assisted prototyping tool (https://www.calicat.cn/). By entering natural‑language prompts, the tool generated five pages: product list, add, edit, delete, and a data‑dashboard, each with an embedded AI assistant. Screenshots of the generated pages are included.

Tool Selection and Initial Trials

Several AI programming tools were evaluated:

OpenAI’s GPT‑5.1 (released 12 Nov)

Google Gemini‑3 (released 18 Nov, claimed to surpass GPT‑5.1)

Anthropic Claude Opus‑4.5 (released 24 Nov)

After encountering authentication issues with Google’s Antigravity, the author switched to domestic tools, ultimately choosing Trae because it was free, generated more professional UI than Tongyi Lingma, and did not charge for large‑model usage.

Development Process with Trae

Using Trae’s Builder mode, the author uploaded the prototype images and provided structured prompts that specified:

Front‑end should use Vue.js; back‑end should be Python with SQLite (easy to replace with MySQL later).

Logging must be enabled.

Python packages should be installed via a mirror to speed up downloads.

Commands should be executed one at a time to avoid Windows incompatibilities.

Initial product data should be pre‑populated.

The AI analyzed the requirements, generated front‑end code, installed dependencies, created back‑end services, set up the database, and integrated a DeepSeek‑based intent‑recognition module for the chat assistant. The process took about 1.5 hours, with the AI iteratively fixing issues such as blank home pages, missing AI assistant widgets, and CRUD functionality bugs.

Key Challenges and Resolutions

Typical problems included:

Empty home page after initial deployment—resolved after multiple AI‑driven debugging cycles.

AI assistant failing to create products, repeatedly returning “product already exists”—identified as a double‑submission bug caused by the confirmation step, fixed by adjusting the dialogue flow.

Incorrect data in the dashboard (e.g., wrong brand names, misplaced charts)—addressed by prompting the AI to generate scripts that query the SQLite database and correct visual elements.

Throughout, the AI provided summary messages after each fix, and occasionally generated temporary scripts to inspect database structures.

Final System Features

The completed application includes:

Product list with pagination and adjustable page size.

CRUD operations for inventory items.

Dynamic dashboards showing total products, items on sale, low‑stock alerts, category distribution, price distribution, and brand distribution.

An AI chat assistant (powered by DeepSeek) that can create products via multi‑turn dialogue.

Code Overview

The project follows a clean front‑back separation. The back‑end contains routes, AI service modules, intent‑handling logic, data models, SQLite files, and initialization scripts. The front‑end consists of Vue components, assets, stylesheets, and routing configuration. The AI service code includes an API‑key configuration for DeepSeek, a prompt template, and a fallback regex‑based intent parser.

Reflections

1. Tool capabilities : Trae’s LLM struggled with multimodal understanding of the prototype images, while Gemini‑3 produced a near‑perfect HTML replica of the UI. However, Gemini‑3 exhibited stability issues when accessed from abroad.

2. Can AI replace programmers? The experiment shows AI can accelerate development of simple modules, but full‑scale, production‑ready systems still require human oversight for bug fixing, performance tuning, and architectural decisions.

3. Embracing AI : The author advocates a “Spec‑Driven Development” (SDD) approach, where clear specifications guide the LLM, improving reliability and reducing hallucinations.

References

[1] AI Native Application Architecture Whitepaper, Alibaba Cloud: https://developer.aliyun.com/ebook/8479