Does AI help more when you start a project from scratch than when you're deep in a mature codebase you already know by heart?

This explores whether AI coding tools help more when you're building something new than when you're working inside a large, long-lived codebase. No study in the collection compares the two directly. The evidence it does have leans toward yes, for reasons that are more interesting than 'big codebases are hard.'

The strongest signal comes from a randomized trial that did not set out to answer this question. Experienced open-source developers worked on real tasks in projects they knew well, and with early-2025 AI tools they finished 19% slower. They had predicted a 24% speedup Do AI coding tools actually speed up experienced developers?. One of the explanations the researchers gave is the surprising part: the developers' deep familiarity with their own mature codebases. When you already hold years of unwritten conventions, edge cases and history in your head, the AI's suggestions are often things you would have done faster yourself, or plausible code that breaks a rule the model never saw. In a mature codebase, much of the important knowledge isn't in the code at all.

That ties to a broader point about what AI actually works from. Its context is whatever happens to be in the prompt, the conversation history and the retrieved files. That context is temporary and changes from turn to turn, unlike the stable picture a long-time maintainer carries around How does AI context differ from conventional software context?. A small greenfield project can fit almost entirely in that window. A mature one can't, so the model is always working from a partial view. When self-improving coding agents were left to evolve, some of the abilities they found on their own were better context management and code editing Can AI systems improve themselves through trial and error?. That hints that handling context is a real bottleneck on serious codebases, not a side issue.

The picture from inside a company is mixed. Anthropic's engineers reported about a 50% productivity boost and merged more pull requests, yet most said they could fully hand off only 0–20% of their work Does AI assistance erode the skills needed to oversee it?. The rest still needed a human who understood the system. So AI helps in established code, but mostly as a collaborator, not something you can hand whole tasks to.

There's also a twist that runs the other way. 'Greenfield' can also mean 'new to you', and there the risk shifts from speed to learning. In a trial where developers picked up an unfamiliar library, AI use hurt their understanding and debugging ability. The exception was developers who stayed actively engaged and asked questions to understand the code Does AI assistance actually harm the way developers learn?. So the setting where AI feels most helpful, unfamiliar territory with few constraints, may be where you build the least of the knowledge that later makes the codebase 'mature' in your own head. The likely answer: AI speeds up greenfield work more, but on fresh projects the cost you need to watch shifts from lost time to lost understanding.