GitHub, VS Code set rules for AI coding tasks

The starting point of this piece is a habit many people fall into: asking an AI to "build the article management feature" and beginning there. That request alone leaves open what gets stored, what happens when the same ID arrives twice, and what counts as finished — and once the code exists, making those decisions from zero is painful. The article's answer is not to write a detailed spec up front, but to write down the scope of this one change and how the finished thing will be checked.

From there, the article draws on two official sources. GitHub's material for its Copilot cloud agent recommends making the problem to solve, the completion conditions, and hints about which files to change explicit. VS Code's documentation describes a mechanism for saving shared development policy in a Markdown instruction file, while noting that the supported format and the way it is read differ depending on the agent chosen. The article's example request shows what this looks like in practice — listing what to save, what to check first, and four things to verify when done, including that re-registering the same ID must not overwrite an existing article.

The remaining thread is verification. The article separates a file merely being found from instructions actually being followed, and suggests trying a representative task under the same conditions and comparing referenced information, generated files and results. It also warns that "tests passed" is not the same as the behavior you cared about being confirmed, and asks for a note on what was tried and what could not be tried. The wider context is that instruction files generated by AI should themselves be reread, since they may describe test procedures that do not exist yet or point at the wrong location. Whether this approach holds up likely depends on how much of the request is genuinely one-off and how much is a rule worth keeping in the repository.