Claude Code and Drupal: Help, Harm, and a Hard Reset

The experiment began in December 2025 with curiosity and cautious optimism. Claude Code produced a working Firefox extension on the first attempt and handled batch media scripts after some back-and-forth. It even helped Théodore generate a mass-update command for GitLab merge requests using an architecture he had already built.

But the wins faded quickly. When asked to refactor code supporting a Drupal bot, the model introduced regressions that went undetected for weeks due to missing test coverage. A credit attribution test—evaluating how well a language model could interpret issue metadata against Drupal core’s credit guidelines—showed promise but remains unfinished. His PHP prototype of the npm CLI demonstrated the feasibility of managing JavaScript dependencies without Node, but Théodore saw its limits. LLM-generated code quickly bloated, added unrelated features, and suffered from low structural discipline unless rigorously prompted.

The second post, titled “…and now I'm recovering”, marks a pivot. Théodore cancelled his Claude Code subscription, citing obsession, sleep loss, and lack of real learning. “Coding agents have a way to turn your brain off,” he writes, calling them the Shein of software: cheap, tempting, and environmentally harmful. He warns that the apparent productivity may come at the cost of judgment and skill.

His critique extends to the broader culture surrounding AI tools: paywalls, hype, and a shift from knowledge to prompt-chaining. Théodore proposes low-tech safeguards, such as adding AGENTS.md files in Drupal core and contrib repositories to help contributors flag AI-generated code responsibly.

In closing, he urges developers to ask harder questions: is the problem worth solving, or is “AI” a crutch masking deeper issues in tooling and team environments? Rather than chasing illusions of speed, he suggests slowing down, talking to peers, and keeping the human layer intact.