What happened

Claude Code’s /code-review command now changes its review strategy based on the selected effort level.

Previously, the command used one fixed review prompt. Anthropic says the rewritten review system now trades speed and token use against broader coverage.

The local command reviews the current branch’s changes and any uncommitted work. It looks for correctness problems alongside opportunities to simplify or reuse code.

You can run /code-review with the session’s current effort setting. You can also pass a level directly, such as /code-review high.

How local effort levels differ

Anthropic describes low effort as a single pass over the changed code. It is intended to be quick and inexpensive enough to run before routine pushes.

Medium effort reads the changes in their surrounding context. It uses several passes to look from different angles, then verifies findings before reporting them.

High effort moves those finding and verification passes into separate agents with fresh context. That separation reduces the chance that a reviewer simply repeats the assumptions of the agent that wrote the code.

Claude Code also supports xhigh and max effort on models that offer those levels. Anthropic describes high through max as broader local review settings that may return less certain findings. The exact effort options depend on the active model.

Higher local settings expand the search beyond the most obvious findings. This matters when a small edit could break another part of the application.

Ultrareview is different from those model effort levels. The /code-review ultra command runs a fleet of reviewers in a remote sandbox and independently reproduces findings before returning them. Anthropic positions it for substantial changes that need more confidence before merging.

Why this matters for solo builders

AI-native builders often use the same agent to write and review a change. That is convenient, but it creates a predictable weakness. The reviewer may inherit the reasoning that produced the mistake.

Fresh-context reviewers create a more independent check. They do not replace testing or human judgment, but they can challenge an implementation before it reaches production.

The effort levels also make review easier to budget. Builders no longer need to choose between skipping review and running the deepest option on every change.

A practical policy could look like this:

  • Use low for copy changes, small UI adjustments, and routine fixes.
  • Use medium for normal features that touch several files.
  • Use high for authentication, payments, data changes, and difficult refactors.
  • Try xhigh or max when a difficult local review justifies more reasoning and token use.
  • Use Ultrareview before merging a large or unusually risky change when remote processing is acceptable.

The category of change matters more than its line count. A five-line permissions edit may deserve more scrutiny than a large generated file.

Ultrareview costs extra and sends code to the cloud

Local /code-review runs inside the current Claude Code session and counts toward normal usage. Ultra sends the repository state to an isolated Anthropic cloud environment.

Anthropic’s Ultrareview documentation says Pro and Max users receive three one-time trial runs. Later reviews typically cost $5 to $25 in usage credits, depending on the change.

The remote review usually takes five to 10 minutes. It is unavailable through Amazon Bedrock, Google Cloud’s Agent Platform, Microsoft Foundry, and organizations using Zero Data Retention. When remote Ultra is unavailable, Claude runs a local review instead.

Builders should check the scope and estimated charge shown before launching it. A repository may contain more than the files changed in the current task.

Treat vendor benchmarks as a starting point

Anthropic says low effort finds more issues than other code-review tools at a lower token cost. The announcement does not publish enough methodology to treat that as an independent comparison.

The safer test is inside your own repository. Record which findings are correct, which are noise, and whether deeper settings catch problems the lower level missed.

Claude’s current Code Review documentation says lower settings return fewer, higher-confidence findings. Higher settings provide broader coverage and may also surface less certain findings.

That trade-off matters. More comments do not automatically mean a better review.

What builders should do next

Update Claude Code and run a low-effort review before your next push. Confirm every finding against the code instead of accepting fixes automatically.

Repeat the review at medium or high on one meaningful change. Compare the additional findings and note how many lead to real corrections.

Reserve Ultra for a change where extra confidence can justify its cost and remote processing. Code review works best as a risk control, not a ritual applied at maximum depth every time.

End of article