Save this and check it later If you use Claude Code Desktop, make sure you have these features enabled You’ll use far fewer tokens and keep your context window much cleaner

虽然我很讨厌A/ 但是不得不说 这个还是很厉害的

Good checklist. Enabling those keeps the context window from filling with noise and cuts the token burn on longer Claude Code runs.

Potente! Gracias Daniel!

De nada 🫡 y se vienen más tips

A cache hit needs the exact same prefix, so the turn you compact is the turn you pay full input rate on whatever survived the compaction. Cheaper across a long session, more expensive right at the cut. Fewer tokens and cheaper aren't the same axis.

The highest-leverage setting is usually the one that makes the model compress state before the context fills. I’d pair these with a lightweight checkpoint habit so a long run can recover without replaying the whole conversation.

Settings like these are per person and per machine. One engineer reads this and flips them, the other nine never see the post, so the team bill moves by a tenth of what it could.

leave every desktop feature on and the window is sludge by message 8. the save is just turning half of that off.

Context window and token count are the easy wins. The one that moved more for us was logging what each tool actually returned, because a tool that returns nothing costs a full turn and looks identical to one that worked. Do you log tool returns or only tool calls?

Worst advice ever... Omg it doesn't work that way!

Yes it helps

how much token usage did those features cut in your longer Claude Code Desktop sessions? a before-and-after on context size would make this tip much easier to act on.

I just applied these settings. The context doesn’t seem to have changed much to me—maybe I’m using fewer tokens like you said, but I don’t see it.

Dirty context has burned more of my budget than a weak model ever did. Which of those Desktop toggles cut your spend the most?

I don't have the informative little circle at the bottom

Good reminder: the best token saving is knowing what already sits in the session and we can get that easily with Token Meter: