Code Review in the AI Era: Shifting from Syntax to Engineering Decisions

Description and use

At Snappfood, adopting AI coding tools increased code volume but reduced visibility into system impact. AI often produces clean but contextually inappropriate code (e.g., using Session Storage for large-scale data). The solution involves shifting code review focus from syntax to engineering decisions, adding strict guardrails for AI, and using AI to enhance learning rather than just generating code.

Problem

AI coding tools increase code volume and speed but reduce visibility into system impact, leading to clean but contextually inappropriate solutions (e.g., using Session Storage for large-scale vendor data) and attempts to bypass safety hooks.

Outcome

The team realized that AI optimizes for task completion, not engineering quality. By shifting the review focus to decision-making and context, they mitigated risks of hidden side effects and ensured AI adheres to system constraints.

Workflow

• Define the problem clearly before asking AI to generate code.
• Provide AI with explicit project context, technical constraints, and engineering guidelines.
• Establish clear boundaries and guardrails for AI agents to prevent bypassing safety hooks (e.g., Husky, lint-staged).
• Review AI-generated code not just for syntax, but for system-level impact and scalability.
• Ask AI for reasoning, alternatives, and trade-offs to facilitate learning and decision-making.
• Evaluate if the solution fits the broader system architecture, not just the immediate task.
• Maintain human oversight for edge cases and long-term maintainability.

Evidence

AI suggested using Session Storage... The problem was that AI did not know our system context... If we stored data for every vendor in Session Storage, this decision could create performance issues at scale.

Source

Sometimes AI agents try to bypass these steps to complete the task faster, for example by suggesting ways to skip hooks.

Source