Your CLAUDE.md is not magic

The author observes that a Claude session can review changes already examined by another Claude session, yet a human remains accountable for the approval process. They note that review has become more hurried due to the increased volume of software production, but depth has struggled to keep pace. The author advocates for checks that run independently of the coding agent, as merely adding an automatic reviewer to a pull request does not guarantee the desired outcomes.

They emphasize the importance of human judgment in evaluating acceptance conditions that require assessment beyond mechanical evaluation. The author stresses that a distinction between requiring a review and its reliability is easy to lose, as running a check does not guarantee the accuracy of its verdict. They argue that both kinds of checks, those that can be evaluated mechanically and those that require human judgement, should happen independently of the coding agent to ensure dependability in the execution of a required model review.

The author also points out that merely adding an automatic reviewer to a pull request does not address the underlying concerns about the acceptance process. They express their frustration with Claude adding excessive code comments, despite specific instructions to use fewer, as it highlights the limitations of instructions as a mechanism for controlling output.

The author questions the value of creating elaborate instruction sequences, as merely stating "always" inside them does not guarantee compliance. Instead, they suggest that the check should run even if the coding agent forgets it, ensuring that critical requirements do not slip through. The author also acknowledges the potential value of using plugins to encode human knowledge and experiences, but warns against producing a large workflow document that merely repeats what another person could have produced by asking the same model.

They emphasize the need for human discussion and decision-making in protecting codebase-specific acceptance criteria, rather than relying solely on an instruction file. The author proposes a team proposal for human stewardship of model-review criteria, recognizing that building an elaborate instruction sequence does not guarantee that the model will follow it.

They advocate for checks that run independently of the coding agent, while acknowledging that the context and human involvement remain essential in ensuring the reliability and quality of the codebase.