The Shift from Generation to Comprehension
Analysis of 116 AI interaction sessions at Sentry revealed that 67% of AI usage was dedicated to comprehension, while only 2% was used for code generation. In a codebase with 15 years of history and 100 daily PRs, the primary challenge is not writing code, but understanding the existing architecture, conventions, and history to ensure changes do not introduce regressions or technical debt.
The 'Catch Me Up' Framework
To standardize this understanding, the author developed a personal skill called 'Catch Me Up.' This is a structured, local prompt-based tool that forces the AI to explore the codebase through six specific modes:
- Architecture: Understanding the high-level structure and component relationships.
- Conventions: Identifying established patterns and linting rules.
- Feature Traces: Mapping how specific features or bugs move through the system.
- Syntax: Clarifying language-specific implementation details.
- Testing: Understanding how to validate changes within the existing test suite.
- History: Contextualizing why specific decisions were made.
Aligning Mental Models Before Implementation
Drawing on the philosophy of 'vibe coding' vs. disciplined engineering, the author argues that AI workflows must include an explicit 'understanding' phase before planning or implementation. By using the 'Catch Me Up' tool to verify the AI's grasp of the codebase first, the developer can steer the agent away from incorrect assumptions. This prevents the creation of 'slop'—code that is generated without context—and ensures that the developer remains the final authority on the codebase's integrity. The AI serves as an tireless, low-cost senior engineer, but it is the developer's responsibility to ensure the mental model is aligned before authorizing any automated changes.