A Non-Developer's Vibe Coding Experience: Leveraging LLMs and AI Agent Tools

A non-developer shared their real-world experience of building and using work-related HTML tools through Vibe Coding. The author realized that building tools directly was faster than searching for them, and compared the usage of LLMs like ChatGPT, Claude, and Gemini with Copilot Agent.

Building Rule-Based Tools with LLMs

LLMs have inherent uncertainty, producing different results for each request, which can reduce reliability for repetitive and rule-based tasks. Consequently, the author adopted the approach of creating rule-based HTML tools for simple repetitive tasks such as text conversion or format correction. For example, they used LLM prompts to generate a 'Keyboard Converter' to restore Hangul incorrectly typed in English, and a web page for real-time editing of Mermaid diagrams. These tools become more reusable when input formats and example data are explicitly displayed on the initial screen.

Utilization and Limitations of AI Agent Tools

Agent tools like Copilot Agent can generate UX-conscious results from a single-line prompt, offering convenience for complex tasks such as navigating between files or fixing errors. However, when reflecting complex requirements, unclear initial prompts can lead to the loss of existing features or results that differ from the intent. Therefore, for complex projects, it was effective to write specific initial prompts and have the LLM draft the prompt, then manually refine the details.

Limitations of Non-Developers and Future Directions

The author pointed out that while AI-generated outputs are useful for creating 'something that works,' they lack the expertise required for actual service operations, such as security, performance management, and test automation. In the future, the author plans to consider adopting terminal-based coding agents like Claude Code or OpenAI Codex CLI, and explore whether to pursue a path of learning development knowledge or adopting an AI-native approach.