My Claude Code PR workflow
How I use AI agents for repo inspection, implementation, verification, and pull requests without skipping engineering judgment.
AI is most useful to me when it makes the work more inspectable. A good agent workflow does not mean handing off judgment. It means tightening the loop between context, implementation, and verification.
The first step is always repo orientation. Before touching code, I want to know the branch, the diff, the relevant files, the existing patterns, and the test surface. If the agent starts by proposing a rewrite before reading the code, the process is already off track.
The second step is a scoped plan. I prefer a short plan that names the files likely to change, the behavior expected to change, and the checks that will prove it. This keeps the work from turning into a broad refactor disguised as a feature.
During implementation, I keep changes small and reviewable. The agent can draft code quickly, but I still read for architecture drift, hidden data assumptions, accessibility regressions, and unnecessary abstractions. The best AI-assisted PRs look boring in the right way: focused diff, clear behavior, clean verification.
Verification is where the workflow earns trust. I run the build, relevant tests, and browser checks for user-facing work. I also ask for a bug-risk review, but I treat that review as a checklist to verify, not as a verdict.
The PR itself should explain what changed, how it was verified, and what risk remains. If a reviewer cannot understand the intent from the PR description and the diff, the agent did not make the team faster. It just moved ambiguity downstream.
The pattern is simple: inspect first, plan narrowly, implement in the local style, verify like a skeptic, and write the PR for the next human.