Use /clear for a new task, /compact only when you actually need the conversation

Claude Code ships with two slash commands that shorten a long conversation: /clear and /compact. /clear throws the conversation away and starts a fresh session in the same working directory, with the repo's CLAUDE.md and your global ~/.claude/CLAUDE.md reloaded but no history. /compact asks Claude to summarise everything said so far into a much shorter version and keeps going from there. Both free up token budget; only one preserves what the agent has learned.

Most people stumble onto the distinction the first time a long session starts to feel slow or repetitive. The instinct is to reach for /compact — it sounds careful, like you're preserving state — but the right question isn't “how full is the context window?” It's “is this conversation still about what I'm doing now?” If the answer is no, /clear is the correct move. The agent is faster, cheaper, and less likely to drift when it isn't dragging ninety minutes of decisions about a different problem behind it.

Task switch — /clear. You finished the bug fix, you're starting a new feature, the two have nothing to do with each other. Compacting the bug-fix conversation just carries forward a summary the new task doesn't need, costs tokens to re-read on every turn, and quietly biases the agent toward the patterns of the old session. Clear it. The repo's CLAUDE.md still loads, your global preferences still apply, and the only thing you lose is the history you didn't want anyway.

Same task, long context — /compact. You're an hour and a half into one investigation, you've read fifteen files, run six commands, made three decisions. The conversation is full of tool output you no longer need but the state of the work — what you're trying to do, what you've ruled out, what's next — still matters. /compact condenses the chatter and keeps the spine. Run /cost before and after; input tokens typically drop sixty to eighty percent, and the agent should still know exactly where you left off.

Same task, but the agent is confused — usually /clear. Counter-intuitive: if the model is repeating itself, contradicting earlier decisions, or treating stale information as current, the answer is rarely “summarise harder.” Compaction carries the confusion forward in compressed form. Clear the session, re-prime with a one-paragraph summary of where you actually got to (in your own words, not the agent's), and start the next leg from a clean state. This costs you a minute of typing and saves you an hour of arguing with a model that's still anchored to a wrong premise three tool calls back.

The trap to avoid is treating /compact as a free token discount. Compaction is lossy. Specific filenames, exact error messages, the precise wording of a decision earlier in the day — those get blurred or dropped. If the next step depends on a fact currently sitting six tool calls back in the transcript, compacting may quietly delete the thing you needed. When the choice is close, decide whether you'd rather lose the conversation's specifics or its momentum, and pick the loss you can recover from with a single re-read.

See the try-it block for the two-keystroke habit that turns this into an automatic call instead of an agonising one.