Every codebase accumulates technical debt. Static analysis tools catch some of it - unused imports, overly complex methods, code style violations. But they miss the debt that requires understanding intent: a function that does three unrelated things, a workaround that was meant to be temporary two years ago, or an abstraction that no longer matches the domain.

The AI Approach

An LLM reads code with an understanding of software engineering principles that goes beyond syntax. It can identify functions that violate single responsibility, spot patterns that suggest copy-paste reuse instead of proper abstraction, and flag code that contradicts its own comments or documentation.

Three-Step Build

  1. Scan - Feed source files to an LLM with a prompt focused on identifying specific debt categories: duplicated logic, dead code, misleading names, overly complex functions, and temporary workarounds that became permanent.
  2. Quantify - For each finding, estimate the severity (how much does this slow down future development?) and the effort to fix. Group findings by area of the codebase.
  3. Prioritize - Rank findings by a combination of severity, fix effort, and how frequently the affected code is changed. Debt in code that nobody touches is less urgent than debt in code that changes every sprint.

Where It Breaks

The AI may flag intentional design decisions as debt. Code that looks like a workaround may exist for a reason the AI cannot infer. Every finding needs human judgment about whether it is actually debt or a deliberate trade-off.

The Production Path

Run monthly against the full codebase or on each PR against changed files. Present findings in a dashboard that tracks debt trends over time. Integrate with sprint planning so the team can allocate capacity for debt reduction based on data rather than intuition.