A new `claude plugin eval` command lets you measure the actual value your plugin is adding. You create test cases, run your plugin against them, score the results, then run the same cases without the plugin to see the delta.
Running `claude plugin eval` scores each test case with and without the plugin in the terminal, and generates a full HTML report. Accounts that support it also get the report published as a private artifact.
Production code written by Claude should clear a higher bar than human-written code, not a lower one. The recommended stack at Anthropic includes lint rules, tests, Claude-driven end-to-end tests, daily fuzzers, and automated code and security reviews. If quality slips, the levers are: use Opus 5 or Fable 5.1, raise effort to high or xhigh, invest in CLAUDE.md, steer more actively, or wait for the next model.
Anthropic's latest Threat Intelligence report examines how the same capabilities that make models useful, like coding ability and biology assistance, also make them dangerous. The report argues safeguards and monitoring are essential as models grow more capable.
The /diff command now opens as a persistent side pane that scrolls, accepts clicks, and updates in real time as code changes. It lets you inspect diffs without switching windows.
A user asked for a way to turn off inline file-write diffs in Claude Code. The answer is /focus.
Claude Desktop ships its own diff viewer separately from Claude Code's new /diff pane. The two products handle this feature independently.
Bcherny endorses the principle that review rigor should scale with risk: throwaway scripts can ship without scrutiny, but anything touching auth or money warrants line-by-line review.
Bcherny reads diffs when code touches something critical or important to get right. He doesn't write the code himself; his job is reviewing what Claude produces in high-stakes areas.
Claude Code's TUI shouldn't be jumping or flickering. The fix is to run claude update first; if the problem persists, run /tui.
The relevant question isn't whether a tool can be misused (anything can) but how much damage it can do when it is. The "blast radius" framing is bcherny's lens for thinking about AI risk.
Responding to a user who accused Anthropic of running a fear campaign to drive AI regulation, bcherny points to the Threat Intelligence report itself as the source of information and encourages readers to form their own conclusions.
A new toggle in the Claude Code desktop app prevents your computer from sleeping during a session, including idle time between turns, so long-running tasks don't get interrupted.
Upgrading the underlying runtime to Bun v1.4.1 reduced idle memory usage for Claude Code on Linux by 14%.
