Stop hitting your token limit

The limit that stops your afternoon is not really a token count. Subscriptions budget compute, and compute spent is tokens consumed times the cost of the model that reads them. A heavy model re-reading a bloated context burns the budget from both ends at once. That formula is the whole strategy: every fix in this guide either shrinks the tokens or downshifts the model, and once you see the two variables you can stop guessing about where your afternoon went.

01Audit before you optimize

You cannot fix a spend you have not seen. In Claude Code, /usage shows how much of your limit is gone and a breakdown of what is using it. Two findings show up for almost everyone: a large share of usage running above 150K context, and a large share coming from subagent-heavy work. Both are token problems, not model problems, so that is where to start.

The second instrument is /context, run twice. Mid-conversation, it shows how fast the window is filling. In a brand-new chat, it shows something more surprising: everything preloaded before you type a word. Tens of thousands of tokens at message zero is common, and that number is your minimum spend for every conversation you will ever start.

02Shrink the tokens

Session habits come first because they cost nothing. Run /clear when you switch tasks instead of letting an old conversation’s context ride along under the new one. Reply while the prompt cache is still warm: the discount expires after a few idle minutes, and a cold cache means the whole context is re-read at full price. Match the effort setting to the task instead of leaving it high for everything. And when the context indicator passes about 60 percent, /compact folds the history into a short summary rather than dragging every earlier turn into every later one.

Then clean up the preload you found in Section 01. Delete MCP servers you do not use; each one loads its tool definitions into every conversation. Archive skills you never invoke and shorten the descriptions of the ones you keep, because names and descriptions load at every session start. Cut CLAUDE.md hardest of all: it is re-read on every message, so a 4,000-token file taxes every exchange. Keep it to how Claude should work with your project, a couple hundred lines at most, with detail split into files loaded on demand. That keep-it-thin habit is the subject of Every new repo starts with your lessons.

Output is the same bill from the other side: every word the model says back to you is billed and then re-read on every following turn. Thinking on a budget covers that half, from a one-line “be concise” to a full compression contract.

We use a skill called caveman for this, on every session behind this site, the same one the video reaches for. It has the model answer like a smart caveman: it strips articles and filler words, allows sentence fragments, swaps causal narration for arrows, and never shortens the things that matter (code, error strings, facts, caveats). The box below shows the top of our version; the button copies every line of it.

---description: Ultra-compressed prose mode. Cuts response token spend by speaking  like a smart caveman while keeping full technical accuracy. Use when the user  says /caveman, asks for terse replies, or a standing instruction enables it.--- # Caveman: compressed prose mode Respond terse, like a smart caveman. All technical substance stays; only fillerwords are cut. Accuracy first, brevity second. Never drop a fact, caveat, or qualifier to savetokens: compress wording, not meaning. If terseness risks a mistake or misread,spend the words. ## Persistence

Run it at whatever intensity suits you; ours stays on ultra.

03Downshift the model

As for the model itself, the rule of thumb is: if AI could solve the task a year ago, it does not need a frontier model today. Scraping, summarizing, formatting, fetching files. Running those on your most expensive model is hiring the law firm’s senior partner to do the photocopying; the partner should lead the case while junior associates carry the grunt work at a fraction of the rate.

The way to apply this without re-deciding every time is to encode it in your skills. A skill definition can pin the model it runs on, so the minimum viable model becomes a one-time decision that every future run inherits. It can also fork its context, starting the task on a fresh thread when it needs nothing from the conversation, which cuts the token side at the same time. The economics of this orchestrator-and-workers split, and how to route by taste as well as cost, are the subject of Rent the model, own the method.

The floor below the cheapest model is no model. A script runs the same way every time, costs zero tokens, and never hallucinates, so any step of a skill that is deterministic (renaming, parsing, moving files, formatting output) belongs in code, with the model kept for the parts that need judgment.

04Riskier fixes, priced honestly

Past the habits and the routing sit heavier fixes, each with a real trade-off attached. Input compression tools such as RTK sit between your commands and the model, deterministically stripping boilerplate and repetition from tool output before Claude reads it; the project claims 60 to 90 percent savings on the text it filters, and the trade is another moving part in your pipeline. Stranger still, there is pxpipe, which renders bulky context as PNG images because dense text packs about three characters per image token against one per text token. It is genuinely clever and genuinely lossy: exact strings can come back garbled, and the pricing quirk it exploits may not survive.

Further out: routing execution-heavy work to a second agent with its own budget, swapping Claude Code’s backing model for a cheaper provider via environment variables, or running an open model on your own hardware. Each buys capacity and sells something you may care about, whether model quality, who controls your data, or your evenings spent maintaining a server. The honest reading of these options is an ordering: exhaust the free fixes in Sections 02 and 03 first, because they attack the same two variables without adding a single new failure mode.