Your Claude Code limit dropped 17%: the 3 tweaks that make it last longer

If your Claude Code limit seems smaller this week, that is not your imagination. The promotion that gave you 50% more weekly limit ended on September 13, and the ceiling that replaced it is lower than what you were using.

The arithmetic is simple: 100 was the original limit, the promotion took it to 150, and what applies now is 125. Twenty-five percent above May, seventeen percent below what you had last week. Both percentages are true because they measure against different baselines, and that is where the catch lives: Anthropic's announcement mentioned only the increase, and the company only published the number users actually feel after the community pushed back.

Complaining about the number does not bring it back. Spending less does, and that comes down to three things I change on my end. All three are free, and the order matters.

1. Run /context before anything else

/context lists what is already loaded before you type your first word: CLAUDE.md, auto memory, MCP tool names, skill descriptions. That goes into every message you send. You pay for that list on each one.

The thing is, nobody reviews that list. A skill from a project you finished in June, an MCP server from a two-week experiment, a CLAUDE.md that grew by accumulation. None of it helps with what you are doing right now, and it keeps getting sent along.

What I do: run /context, see what is heavy and cut what I no longer use. It is not a weekly chore. You do it once and the savings apply to every message after that. It has the best payoff of the three because it is the only one that needs no habit afterwards.

2. caveman: same answer, fewer words

A long answer costs you twice. Every output token is spent right away, and on the next message that answer comes back as input. Decorative prose is the expensive part of both.

caveman is a plugin that makes the agent answer in short form. Same diagnosis, same fix, no throat-clearing. Code, commands, file paths and error messages are never trimmed, only the prose around them. I use it here every day.

The gain is real, but smaller than the number going around. The author's benchmark measures output only: ten ordinary coding tasks went from 1,214 to 294 response tokens on average, 65% less. Across a whole session the discount is much smaller, because the plugin costs 1 to 1.5k input tokens per turn and does not change what the agent reads or how much it reasons. The author himself warns that on work that was already terse it can break even. Worth knowing before you bet everything on it.

To install:

npx skills add JuliusBrussee/caveman

3. /clear between tasks

Old conversation does not go away when you switch topics: it gets re-read on every new message in that session. Finished the bug and moved to another document? The entire bug discussion still comes along, pushing out the file you need to read next.

The official docs are direct about this: run /clear when you switch to unrelated work. If the new task is long and relates to the previous one, /compact works better and takes a focus (/compact focus on the auth bug keeps what you chose instead of wiping it).

If your limit already ran out

There is no trick that gives you capacity back: you wait for the reset and stop repeating the spend. When you are back, this is the order I use: /context first, because it is the bigger cut and the only one you do once; caveman after that; /clear as a habit, no ceremony.

Two Claude Code features belong in the same accounting if you want to go past these three: subagents, which do the big reads in a separate context window, and /autocompact, which sets how full the context gets before automatic compaction runs.

Closing

The three tweaks do not bring the 50% back. Nothing does: it was a promotion, and promotions end even when they get extended a few times. What you can do is stop spending limit on things you are not using. Run /context and tell me what showed up. If you get stuck on a step, tag me and I'll help.