/code-review now supports effort levels, each with a fully rewritten review strategy. Low effort already beats other code review tools on findings at a fraction of the token cost; high effort delivers significantly higher recall for deeper dives.
The ultra tier of /code-review goes beyond x-high by spawning a fleet of reviewer agents that independently reproduce every finding, delivering the same severe-issue coverage as high effort but with far fewer false positives. Anthropic uses it on every internal PR.
Previously /code-review ran a single fixed prompt regardless of effort. Now each level maps to a distinct review strategy, and the command picks up whatever effort level is set for the session. You can also override it explicitly, e.g. "/code-review high".
Low effort runs a single pass over the diff, cheap enough to use before every push. Medium reads changed code in context, runs multiple finder passes from different angles, then verifies every finding before surfacing it.
The recommended Claude prompting architecture is three-part: keep the system prompt lean, put the real weight into artifacts and context, and keep skills small too. It's a tidy heuristic for anyone building on Claude.
In a thread on prompt architecture, the guidance is that the ideal pattern depends on what skills actually do. Having lots of scripts in skills is good, and progressive disclosure works well, but the prevailing mistake is that most skill.md files are simply too large.
Responding to a user who wanted skills to be installed without consuming context, the clarification is that skills already work this way: the model only sees the skill's name and its trigger condition by default, loading the rest dynamically.
A user noted that models are noticeably better at calling skills on their own, meaning progressive disclosure of capabilities now works in practice where it previously fell flat. The team member agreed.
Claude is being used to monitor automated crash reports and generate fixes autonomously, but the workflow has a real gap: there's no clear owner for the resulting PRs, which makes reviewing, following up, and merging them awkward in practice.
The Bun repo's pull request queue now features CodeRabbit and a Claude review bot holding full conversations with Robobun, the automated merge bot. The prediction is that within months this chain collapses into full auto-merge: a second Claude instance with a fresh context window judges blast radius, and if it's low, the PR merges itself.
The Bun team could already hand off merge decisions to an automated Claude-based reviewer, but CI and code review rules aren't tight enough yet to trust it. The hesitation is self-imposed: the infrastructure is ready, the confidence in the guardrails isn't.
Both the 5-hour and weekly rate limits have been reset across all user tiers, giving everyone a fresh slate of usage capacity.
Following the rate limit reset for all users, a team member expressed enthusiasm, apparently using the extra headroom to keep working on something called Fable.
High effort runs finder and verifier agents as subagents with fresh context so they aren't anchored on the writing agent's reasoning. X-high goes further, sweeping for impacts to code outside the changed diff.
The rollout is live: update Claude Code and run /code-review on any PR to access all effort levels, including ultra.
