Software engineer Charles-Axel Dein published a post on March 13, 2026 arguing that storing documentation alongside code in git repositories matters more now than it did before AI agents. Dein maintains the charlax/professional-programming GitHub repository and observes that AI agents have already shifted developer behavior: the proportion of markdown files in commits has risen sharply, driven by agentic implementation logs and the proliferation of <a href="/news/2026-03-14-8-levels-agentic-engineering-framework">rules files such as Cursor's .mdc format</a>. His core argument is that these rules files are effectively undocumented style guides — best practices engineers never formally wrote down — and that the line between "AI-only" and "human-only" markdown is blurring to the point where structured documentation may eventually replace rules files entirely.
On the "stale docs" objection — the perennial reason teams skip documentation entirely — Dein argues agents flip the equation. Automated code-documentation alignment checks running during pull requests, or via dedicated review agents, remove the manual overhead that made maintaining docs feel futile. The enforcement becomes part of the pipeline. In-repo docs — RFCs, PRDs, architectural decision records — also give agents higher-level context that reduces the token-intensive exploratory work required to understand a large codebase.
The most novel idea in the post is what Dein calls "metaplans": documents stored in the repo that capture research findings rather than just describing existing system behavior. He spent two weeks iterating on structured logging best practices, then encoded the conclusions as a metaplan reusable by teammates and agents alike. Metaplans sit outside the usual doc taxonomy — ADRs and RFCs record decisions and proposals; metaplans capture the research that produced the decision, the operational knowledge no agent can infer from source code alone. Dein points to Diátaxis, Sphinx autodoc, Docusaurus, and Python's doctest as infrastructure for making this work at scale. For teams, the practical shift is <a href="/news/2026-03-15-developer-builds-cutlet-language-with-claude-code-without-reading-code">treating docs as inputs rather than outputs</a> — something written before the codebase drifts, actively queried by agents rather than regenerated from scratch on every run.