Context window

A context window is the total amount of text a model can hold in mind during a single pass, measured in tokens. In a Claude Code session it includes everything in play at once: your current message, the earlier back-and-forth, every file Claude has opened, and every result a tool has returned. The model cannot reason about anything outside that window, so the context window is effectively the working memory for the task at hand.

Claude models ship with large context windows, but a real codebase fills them faster than people expect. Reading a handful of big files, running a few commands that return long output, and carrying a long conversation can quietly consume most of the budget. When context gets crowded, quality drops: the model has to weigh too much at once, and important details from early in the session can get crowded out by noise from later. Recognizing the symptoms - vaguer answers, forgotten instructions - is the first step to fixing them.

Managing context is a core Claude Code skill. Put standing facts in CLAUDE.md so they load efficiently instead of being re-explained. Open only the files relevant to the current task rather than pointing Claude at the whole repo. Delegate sprawling exploration to a subagent so its reading stays out of your main thread. And when the task changes meaningfully, start a fresh session rather than dragging an exhausted context along. Deliberate context management is one of the clearest dividing lines between people who get reliable results and people who fight the tool.