Lawrence Jengar
Apr 16, 2026 00:53
Anthropic releases /utilization command and context administration information for Claude Code, serving to builders navigate 1 million token classes extra successfully.
Anthropic dropped a brand new /utilization command for Claude Code on April 15, alongside detailed steerage on managing classes throughout the AI coding assistant’s a million token context window. The replace addresses a rising ache level builders have flagged: determining when to proceed, compact, or nuke a session fully.
The corporate’s technical workers member Thariq Shihipar laid out the logic in a weblog publish that reads extra like tribal data than typical documentation. The core downside? Context rot—the place mannequin efficiency degrades as extra tokens pile up and a spotlight will get unfold skinny throughout irrelevant cruft.
5 Methods to Deal with Context Bloat
Claude Code now provides builders a call tree at each flip:
Proceed works when the whole lot within the window nonetheless issues. /rewind (or double-tap Esc) allows you to bounce again to any earlier message and reprompt, dropping the whole lot after that time. Anthropic suggests this beats the intuition to sort “that did not work, strive X as a substitute”—as a substitute, rewind to only after helpful file reads and provides cleaner directions.
/compact asks Claude to summarize the session and substitute historical past with that abstract. It is lossy however automated. You possibly can steer it: “/compact deal with the auth refactor, drop the check debugging.” The catch? Autocompact fires when context is sort of full, exactly when the mannequin performs worst because of context rot.
/clear places you in management. You write what issues and begin contemporary. Extra work, cleaner slate.
Subagents deal with the messiest situation—when the following chunk of labor will generate tons of intermediate output you will not want once more. Spawning a subagent offers it a contemporary context window; solely the ultimate report comes again to the guardian session.
The Psychological Mannequin
Anthropic’s inner check: “Will I want this device output once more, or simply the conclusion?” If simply the conclusion, spin up a subagent for codebase searches, verification runs, or doc writing.
The steerage suggests beginning contemporary classes for genuinely new duties, regardless that 1M context technically permits marathon classes. For associated follow-up work—like documenting a function you simply constructed—staying in session avoids costly file rereads.
Unhealthy autocompacts sometimes occur when Claude cannot predict the place your work is heading. A protracted debugging session will get summarized, you then ask about “that different warning in bar.ts” that obtained dropped from the abstract. With the bigger context window, Anthropic recommends proactive /compact with specific course earlier than the automated set off fires.
The replace displays Anthropic’s push to make Claude Code extra sensible for prolonged improvement classes. Whether or not this strikes the needle on enterprise adoption stays to be seen, however builders burning by context on advanced codebases lastly have documented methods past trial and error.
Picture supply: Shutterstock
