Doc code fences should be assembled at build time, not generated by models

On 2026-11-10, DEV Community's AI channel published an article titled "Documentation Code Fences Should Be Spliced by the Build, Not Generated by the Model." Its central claim is that a documentation pipeline should not let a model write code blocks directly. The reasoning is straightforward: if a model can type a fence, it can retype an example from memory, and the code the reader copies will drift from the function the tests exercise. The author therefore argues that code fences already have an owner, namely the repository, and that a model should receive only the sentences between the code blocks and nothing else.

The workflow described makes fences un-writable by the model. Fences are spliced from marked source regions, and any draft that tries to introduce one is rejected. The goal is to keep documentation examples tied to the code the repository actually ships, rather than to whatever a model happens to produce on a given run.

One important caveat the article stresses: the scripts shown in the piece are illustrative only and were not actually executed. Readers are therefore meant to treat the workflow as a design sketch to adapt, not as ready-to-run tooling. The article does not include benchmark numbers or performance claims.

As of the time of this report, the piece has not been followed up by further articles in this dataset, so the current state is a single published design proposal with no verified implementation or independent replication.