Non-Programmer Ships a Product in 16 Days Using AI Tools: A Complete Toolchain Retrospective

How a Non-Coder Built and Shipped a Product Using AI

An indie developer (self-described as unable to write code) shared their complete AI tool journey from zero to product launch on Bilibili. The entire process took about 16 days, during which they cycled through multiple AI coding tools and large language models, ultimately completing the final deployment with DeepSeek's help. This experience vividly illustrates both the capabilities and real pain points of current AI-assisted programming.

Getting Started: From Gemini Chat to Firebase Studio

Requirement Validation: Using Gemini to Confirm Feasibility

The author initially used Google's Gemini for requirement validation—when they had a product idea, they first asked AI whether it was technically feasible. Gemini is Google's multimodal large language model series, released in late 2023 and positioned as a flagship competitor to OpenAI's GPT-4. It offers text comprehension, code generation, logical reasoning, and more. However, when used through a pure chat interface, the generated code requires users to manually copy it into a development environment to run and debug, lacking an immediate feedback loop. For someone who doesn't understand code, this is a massive barrier—even if the AI generates correct code, the user doesn't know where to put it or how to run it. Since Gemini as a pure chat tool couldn't directly produce a runnable code environment, the author turned to Google's IDE tool "反重力" (literally "Anti-Gravity," referring to Firebase Studio, formerly known as Project IDX and other integrated development environments).

First Wow Moment: AI Auto-Generates Runnable Code

"Anti-Gravity" is actually Google's Firebase Studio, a cloud-based integrated development environment (IDE). Its core advantage lies in deeply integrating AI code generation with a runtime environment—users don't need to configure any local development environment, and AI-generated code can be previewed and run directly in the cloud. Firebase Studio embeds multiple models; the ones the author used include:

• Gemini Pro and Gemini Flash
• Claude Opus 4.0 and its lower-tier versions

Claude Opus 4.0 is Anthropic's flagship coding model, particularly outstanding in code generation and complex logical reasoning, widely regarded as one of the strongest programming assistance models currently available.

The author described their first experience as "mind-blowing"—just state a requirement, the AI automatically generates code, click run, and the functionality matches expectations perfectly. This "what you think is what you get" experience is undoubtedly revolutionary for non-programmers.

Mid-Development: Model Quota Anxiety and Quality Gaps

Good Models Run Out Fast, Weak Models Can't Get the Job Done

However, going from "works for me" to "works for others" to "ready for production" involves completely different levels of workload and bug complexity. The author quickly hit a core pain point: good models (like Gemini Pro) burn through quota extremely fast, often exhausted after just two conversation rounds, forcing a switch to the weaker Flash model.

The difference between Gemini Pro and Gemini Flash comes down to model parameter scale and reasoning depth. The Pro version has more parameters and stronger reasoning capabilities, but each call consumes more computational resources, so the platform imposes strict usage quotas. The Flash version is a lightweight, fast-response model—quicker and cheaper, but noticeably weaker at handling complex logic, understanding long contexts, and multi-step tasks. This "premium models in limited supply" business model fundamentally reflects the industry reality that LLM inference costs remain high—every high-quality model call means real GPU compute costs paid in real money.

The experience gap was stark:

• Good models: one sentence gets the job done
• Weak models: ten sentences might not solve it, and they not only fail to understand instructions but also mess up previously working code

This reflects a universal problem with current AI coding tools—the massive contradiction between high-quality model usage costs/quota limits and actual development needs.

Free-Tier Strategy: Exploiting New User Credits

To get more high-quality model usage, the author registered new accounts across multiple IDE tools, leveraging free trial credits to complete feature development. Tools mentioned include:

• Kilo (AI coding IDE)
• Windsurf (an AI coding product from the same era as Cursor)

Windsurf, developed by the Codeium team, focuses on an "AI Flow" programming experience that proactively understands developer intent and executes multi-step operations continuously. These products typically use a "free trial + subscription" business model, giving new users a certain amount of high-quality model calls upon registration (usually worth $20-50), designed to let users experience the best results before converting to paid subscribers. The author essentially converted multiple platforms' marketing budgets into their own development resources.

These tools' free credits typically come paired with high-quality models like Claude 4.6, helping the author complete some critical feature development.

Late Stage Workhorse: The Cline + MiniMax Combination

The Idea: Put Idle Token Resources to Work

The author had previously purchased a MiniMax annual token package for using "小龙虾" (a general AI assistant tool), with monthly quotas going unused. MiniMax is a Chinese AI company whose large model performs well in Chinese comprehension and general conversation, but has relatively limited capabilities in specialized domains like code generation and system operations. So the author started thinking about how to utilize these idle resources—like paying for a data plan but barely using the internet.

Cline Plugin: Giving Your IDE a Swappable AI Brain

They found Cline, a plugin that integrates into VSCode or Firebase Studio, essentially giving the coding tool a user-selectable AI brain. Cline is an open-source VSCode plugin (formerly Claude Dev) whose core design philosophy is "model agnosticism"—it doesn't bind to any specific AI provider but connects to any large language model through standard API interfaces. Users just need to provide an API Key and model endpoint address to plug any OpenAI API-compatible model into their coding workflow. Unlike closed AI IDEs like Cursor and Windsurf, Cline gives users complete freedom of model choice, meaning they can flexibly switch between different models based on task complexity, cost budget, and response speed.

By connecting MiniMax through Cline, the author achieved a stable development experience.

Why General AI Assistants Aren't Suited for Coding

The author specifically explained why general AI assistants like 小龙虾 can't be used for coding: programming requires a holistic view of the entire project and understanding of context, areas where chat-style assistants are inherently weak.

This involves a core technical concept in AI programming: the Context Window. Professional AI coding tools package the entire project's file structure, relevant code files, dependency relationships, and other information into the model's context, allowing the AI to "see" the complete project picture. General chat assistants can only process code snippets that users paste in each conversation—they can't automatically index project files, understand inter-module call relationships, or directly modify code in the file system. It's like asking a doctor to diagnose a condition by looking at only a cropped section of an X-ray—lacking global information inevitably leads to misjudgment. This is the core value proposition of professional AI coding IDEs.

The MiniMax model helped the author resolve numerous small bugs and feature adjustments during the sprint phase, but also exposed shortcomings:

• Struggled with complex tasks like server deployment
• Extremely slow responses when context grew too long (waiting 30 minutes for a reply)
• Suboptimal response quality

The Final Push: DeepSeek Solves All Blockers in Five Minutes

In the final stage before launch, the author was stuck on several stubborn bugs that MiniMax simply couldn't resolve. It wasn't until they switched Cline's model to DeepSeek that all blockers were resolved within five minutes, and the product went live immediately after.

DeepSeek is a large language model developed by China's DeepSeek company. Its V3 and R1 series demonstrate exceptional programming capabilities, repeatedly achieving leading scores in international coding benchmarks. DeepSeek's technical highlights include using Mixture of Experts (MoE) architecture to reduce inference costs and significantly improving code reasoning through reinforcement learning. Server deployment tasks demand more from AI models because they involve knowledge spanning multiple intersecting domains—OS configuration, network settings, environment variables, permission management, process supervision—requiring the model to possess cross-domain systematic thinking rather than pure code generation ability.

This comparison is highly compelling: the same problem, the same tool (Cline), and merely swapping the underlying model produced a qualitative leap. This demonstrates that in AI-assisted programming, model capability remains the decisive factor—tools are merely the vehicle.

Complete Toolchain Review and Key Takeaways

Tool Selection by Phase

Phase	Tool/Model	Purpose
Requirement Validation	Gemini	Confirm feasibility
Early Development	Firebase Studio (Gemini Pro/Flash, Claude Opus)	Core feature development
Mid Development	Kilo, Windsurf	Leverage new user credits for feature development
Late Development	Cline + MiniMax	Daily bug fixes and feature adjustments
Deployment & Launch	Cline + DeepSeek	Resolve final deployment challenges

Three Key Insights

1. AI coding has made it possible for non-programmers to independently develop products, but the gap from demo to production remains enormous. This gap mainly manifests in error handling, edge cases, security, performance optimization, and deployment operations—the "last mile" problems that are precisely the most time-consuming parts of traditional software engineering.
2. Model quality differences are massive—the efficiency gap between good and weak models can exceed 10x. This difference shows not only in code correctness rates but also in the ability to understand ambiguous requirements, grasp project context, and avoid "hallucinations" (generating code that looks reasonable but is actually wrong).
3. The ability to flexibly switch models matters greatly—"model-agnostic" plugins like Cline provide tremendous flexibility. As the LLM market iterates rapidly, today's strongest model may be surpassed within months, so tool architectures that don't lock into specific vendors hold stronger long-term value.

This creator's experience proves that one person leveraging AI tools can indeed complete a product's development and launch within half a month. But the process is far from smooth sailing—it's more like a real-world battle of maneuvering between different AI tools, adapting on the fly.