ground-first

To analyze code behavior before proposing edits, use a grounded analysis approach that reads all relevant files and cites exact line references for every claim. This prevents ungrounded code changes by requiring a full investigation of current functionality before any modifications are suggested.

Incorrect code fixes happen when analysis skips understanding existing behavior. A ground-first approach solves this by mandating a full review of current code behavior, explicitly listing unconfirmed unknowns, and halting edit proposals until you validate the analysis matches reality.

To debug complex code without regressions, mandate a full behavior analysis of existing code before any fixes. This involves mapping out relevant logic and entry points with precise file and line citations, and separating unconfirmed unknowns from substantiated assertions before proposing changes.

Yes, you can validate code investigation results by prohibiting mutation tool use until you manually confirm the analysis. This structured approach requires the AI to cite actual file and line references for all claims, ensuring you confirm understanding of current functionality before any modifications.

The best way to review code changes for non-trivial codebases is to enforce a grounded analysis phase first. This requires using Read, Grep, and Glob tools to locate and review all relevant code, producing a structured analysis with precise citations for every claim before any edits are proposed.