From Vibe to Structure: A Pre-Development Framework for AI-Assisted Software Engineering

Abstract

New AI tools such as ChatGPT and Claude let people build software just by describing what they want in plain language. This way of working is called "vibe coding." A person explains an idea, and the AI writes the code. It is fast, and it is easy to start. It is also very informal. Most developers do not prepare any information before asking the AI to build something. That habit often produces code that misses requirements, contains errors, or does not match what was actually intended. This paper asks a simple question: what should a developer prepare before asking an AI to build software? We propose the Pre-Development AI Context Framework (PDACF), a six-phase framework that guides a developer through project understanding, requirement structuring, system structuring, AI context preparation, AI implementation, and validation, all before any code is generated. The framework builds on established software engineering practice and adapts it to how people actually work with AI coding tools today. To check whether this framework holds up in practice, we surveyed 25 respondents (18 final-year software engineering students and 7 industry professionals) who use AI coding assistants regularly. Each respondent rated their own preparation habits and development outcomes and also reported objective figures such as time spent, number of prompts used, and number of code re-generations. A very strong positive correlation (r = 0.940, p < .001) emerged between how much a respondent prepared before using AI and how satisfied they were with the outcome. Respondents who prepared well finished tasks 49% faster (1.58 hours versus 3.12 hours). They used 53% fewer prompts and needed 65% fewer code re-generations than respondents who prepared little, a difference that a t-test confirmed was unlikely to be due to chance. These findings are correlational rather than causal, and the sample is small, so the results should be read as a first step rather than a final answer. Even so, they give useful evidence that structured preparation, not only better prompting, deserves closer attention in AI-assisted software development.