Encoding Team Standards into AI Instructions
Why senior engineers' tacit knowledge causes inconsistency in AI-assisted development, and how it can be turned into executable, versioned instructions.
On experienced engineering teams, a senior's instinctive judgment about code review, refactoring, and security checks rarely gets written down—it stays as tacit knowledge in their head. Even when junior developers use the same codebase and the same AI tool, they get materially different results, because the gap isn't in what the AI knows about the project but in the quality of instructions it receives. This turns seniors into bottlenecks, not because they write the code, but because they're the only ones who know what to ask for.
The piece argues this is a systems problem rather than a skills problem, and proposes encoding team standards as versioned, executable AI instruction sets rather than wiki pages or oral tradition. Like linting configs or CI/CD pipelines, these instructions should function as infrastructure, not documentation—reviewed via pull requests and living in the repository so standards get applied automatically rather than depending on memory or discipline.
A well-formed instruction has four consistent parts: a role definition (setting the AI's expertise lens), context requirements (making dependencies explicit), categorized standards (encoding priority—must-follow, should-follow, nice-to-have—which captures the team's judgment, not just knowledge), and a structured output format (summary, categorized findings, next steps). This anatomy gives engineering teams a scalable way to make AI-assisted development consistent regardless of who's doing the prompting.