Ankit Jain separates code review into checking correctness and maintaining team alignment. Automated checks can help find bugs and enforce conventions, but mentorship, architectural discussion and knowledge sharing still need a place in the workflow. He argues that increasing AI-generated code makes it more important to redesign review without discarding those human functions.
Ankit Jain explains that specifications alone miss decisions discovered during implementation. Tickets, product requirements and conversations with a coding agent all contain intent, yet prompt history is often discarded when a pull request is created. He proposes turning the human decisions in those sessions into acceptance criteriaContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task. so reviewers can understand what was requested, clarified and rejected.
Ankit Jain introduces an AI slop registry as a maintained collection of recurring review findings and reusable rules. Combining those rules with acceptance criteria produces a test plan that can guide verification in an application previewAgent completion verification checks evidence that an agent achieved the requested outcome rather than merely producing intermediate work.. The proposed review material becomes the intended behavior and collected evidence rather than only a code diff.
Ankit Jain describes a hybrid verification approach: deterministic checksDeterministic software verification checks an output with a repeatable procedure that produces the same result for the same inputs and state. where possible, with agents or LLM judgments where necessary. A payment-form example uses browser interactions, screenshots and database snapshots to assess whether the intended behavior occurred. He warns against deriving the entire test plan from the generated code, because that can repeat the implementation's mistaken assumptions.
Ankit Jain argues that reviewers should assess intent, architectural choices and verification resultsHuman judgment in AI is the accountable interpretation and decision-making people contribute when setting goals, evaluating evidence, and managing consequences., with test plans written clearly enough for product managers and designers to participate. Building the reusable rule registry takes initial effort, but each captured pattern can reduce repeated review commentsAn AI feedback loop occurs when AI outputs or their consequences become inputs that influence later model or agent behavior. while preserving team collaboration.
Watch on YouTube
