Moving Code Review from Diff-Based to Intent-Based Verification

AI Engineer2026-08-17go watch the original →

the gist

Code review is failing because humans are skimming AI-generated code they don't read. The solution is to stop reviewing diffs and start reviewing intent, using captured session data to generate acceptance criteria and automated test plans.

The Shift from Diff to Intent

Code review has become a bottleneck where developers spend 4x more time waiting for reviews than in the past, leading to 30% of changes being merged without any human oversight. Because AI agents now write and review code, human reviewers often perform perfunctory "skimming" that fails to catch semantic errors or provide meaningful architectural alignment. The goal is to move the review surface away from the diff and toward the intent, treating the development session as the primary source of truth rather than the final code output.

Codifying Knowledge with an AI Slop Registry

To maintain semantic accuracy without manual line-by-line review, teams should build an "AI slop registry" that codifies recurring feedback. By mining the last 1,000 pull request comments, teams can identify repeatable patterns and turn them into automated guardrails. This registry acts as a persistent training layer that prevents the same mistakes from recurring, effectively automating the "nitpick" portion of code reviews so humans can focus on high-level architectural alignment.

Generating Evidence-Based Verification

Instead of relying on static specs that become outdated, teams should capture the interactive session between the developer and the AI agent to generate acceptance criteria. This process follows a specific workflow:

Capture the user decisions and back-and-forth prompts from the coding session.
Convert these decisions into structured acceptance criteria using an LLM.
Combine the criteria with the AI slop registry to generate an automated test plan.
Run the test plan against a live preview environment to produce verifiable evidence of success.

This approach shifts the reviewer's role from checking syntax to verifying that the implementation meets the defined intent and behavioral requirements. By using screenshots, database snapshots, and end-to-end tests as evidence, reviewers can make informed decisions about architectural integrity without needing to parse the underlying diff.