Code review friction often stems not from logic errors, but from semantic mismatches between human intuition and LLM-generated naming conventions. Andrew Moffat, writing on his personal blog on October 1, 2026, outlines a specific post-processing workflow to mitigate this cognitive load. The core issue is that AI models frequently invent bespoke terms for abstractions—such as 'MutationIntent'—that do not align with a developer's mental model, which might prefer 'EditRequest.' This mismatch forces reviewers to perform constant mental lookups, accumulating friction across dozens of new terms.

The Naming Friction Problem

Moffat argues that while individual naming discrepancies seem minor, they compound rapidly in large AI-generated changesets. The human brain has a finite capacity for holding these semantic mappings in working memory. When an LLM introduces unfamiliar terminology for abstract objects, processes, and concepts, the reviewer’s ability to accurately interpret system behavior degrades. This is not a bug in the code logic, but a bug in the communication channel between the AI agent and the human architect.

A Prompt-Driven Post-Processing Step

To solve this, Moffat proposes a pre-review cleanup phase using a specific prompt. The instruction asks the LLM to scan the changes and extract all unconventional or bespoke terms. The AI then generates a temporary markdown file listing each term, its meaning, the rationale behind its choice, and proposed alternatives. This step transforms the review process from a passive reading of foreign terminology into an active alignment of vocabulary. The developer reviews this file to confirm or replace terms with their own preferred nomenclature.

Automated Refactoring for Consistency

Once the developer finalizes the terminology choices, the AI executes a global find-and-replace operation across the codebase. Crucially, this refactoring extends beyond just the source code to include documentation, ensuring consistency across the entire project artifact. Moffat notes that the resulting code is significantly easier to review because the abstraction names now mirror the developer’s own linguistic patterns. This effectively bridges the gap between machine-generated structure and human-readable intent.

Key Takeaways

  • Cognitive load in AI-assisted coding often comes from naming mismatches, not logic errors.
  • A dedicated prompt step can extract and propose renames for AI-invented abstractions before review.
  • Global find-and-replace operations must include documentation to maintain project consistency.
  • Aligning AI terminology with developer intuition improves the speed and accuracy of code reviews.

The Bottom Line

This is a pragmatic, low-friction workflow improvement for teams heavily relying on AI pair programming. By treating naming conventions as a distinct, addressable problem rather than a minor annoyance, developers can reclaim significant mental bandwidth during the review process.