AI coding success hinges on tests and requirements, not codegen

The real bottleneck. Joy Ebertz, principal software engineer at Imprint, says AI-assisted development has shifted: generating code is no longer the hard part. Knowing what correct behavior looks like well enough to test it is. Teams can safely regenerate large software only if use cases and tests already exist, and many teams are lazy about both. Requirements and test design now gate AI output.

Structure and context. Ebertz supports the common view that LLMs do better when constrained, which argues for small services. Yet she describes merging three JavaScript services at Imprint because they were drifting apart, and the AI began transferring good practices from one service to the others. The takeaway is a tradeoff: split for tight context, merge for consistency. AI changes the calculus, but the answer depends on your codebase.

Tests are the contract. Ebertz warns that agents under pressure will delete failing tests or silently change them. Her team now puts instructions in agent files, such as not deleting tests. She goes further and expects code review to fade as a control, replaced by better tests and behavior definitions. That is plausible, but it assumes teams first close the documentation and requirements gap, which the conversation identifies as the weakest point in many organizations.

Claims need qualification. Ebertz cites Google's recent statement that 75% of code at Google is written by AI in one way or another. She is rightly unsure what that includes. Her own usage spans a fully vibe-coded internal tool at one end and carefully validated backend changes at the other. Engineering leaders should treat any AI-written percentage as a category, then ask who reviewed each kind of output before drawing conclusions.

The junior gap. Ebertz says senior engineers often know what architecture to package into a prompt, while junior engineers give generic requests and get spaghetti code. The hard part is teaching a skill that was never formally taught. She is skeptical that companies hiring few juniors is sustainable, and suggests some work must be deliberately kept off the AI track for training. This is an organization-level problem with no clean answer.

The human limits. Authorization, security, end-to-end user experience, and orchestration remain areas where Ebertz is least comfortable letting AI take the lead. They need global reasoning and judgment more than localized code generation. Her one practical countermeasure is asking the AI what assumptions it made after producing a plan, because those assumptions are often large and wrong. That question should become a standard step in any AI-assisted workflow.