The bill arrives, and a lot of it was you saying the same thing again.

Everything loads, whether the task needs it or not

The vendor documentation is plain about the cost of the instruction file:

"CLAUDE.md files are loaded into the context window at the start of every session, consuming tokens alongside your conversation."

code.claude.com, How Claude remembers your project

Every session. In full. Including the four hundred lines about the payment flow while the agent renames a variable. That is not waste caused by anyone doing anything wrong. It is what a file loaded whole costs.

Corrections live in conversations that end

You explain a preference. The session ends. Next week you explain it again, and that explanation is billed again, along with the back-and-forth that follows it. The expensive part is not the sentence. It is the ten messages of getting there.

Work reported done is paid for twice

The most expensive tokens are the ones spent building something that was never really finished, because you pay again to find out and again to fix it. The forty-dollar session above is the visible version. The ordinary version is a Tuesday spent redoing Monday.

Four changes, free, in order of how much they return.

1. Look at where it actually goes first

Claude Code has a command that shows what is filling the context window, broken down by category. Run it before your first question of the day, not after the bill. Most people are surprised by which line is the big one, and it is rarely the one they were worried about.

2. Cut the instruction file, then scope what is left

The documented target is under 200 lines, and the stated reason is that longer files consume more context and reduce adherence. So the same edit that saves tokens also makes the rules work better. Move anything that applies to one area into a path-scoped rule, which loads only when a matching file is read.

3. Clear between unrelated tasks

An old conversation about something else is being carried into every message you send now. Clearing between unrelated pieces of work is the cheapest habit on this list and almost nobody does it.

4. Ask for the command before you believe done

This is the one that saves real money rather than pennies. When something is reported fixed or tested, ask what ran and what came back. Thirty seconds there against a day of redoing it later.

Mission HQ addresses two of the four, the two that repeat. It does nothing about the price of a token and it is not a billing tool.

The context arrives scoped to the task

Conventions and past lessons are stored per area of work and delivered when a task of that kind starts. The database rules arrive with the database task. They are not in the window during the button work, so you are not paying to carry them there.

A correction is paid for once

You explain a preference once. It is stored, and it arrives with the next task that it governs, and every one after that. The ten messages of getting there happen a single time instead of monthly.

A completion that must show proof

A task does not close on the agent's word. A test counts when a runner on your machine actually ran the command and posted back the exit code. That is aimed squarely at the forty-dollar session, where the expensive part was believing the report.

A page about somebody else's token bill that hides its own overhead is not worth reading, so here is ours.

Mission HQ ADDS context. When a task starts, the agent loads that task's detail, and that payload is budgeted at about 6,000 tokens: the acceptance criteria, the rules for that kind of work, the module's documented knowledge, and who the work serves. There are also three extra tool calls per task, one to create it, one to load it, one to close it.

So this is not a discount. It is a trade. You pay a known amount per task, scoped to that task, instead of an unknown amount per session carrying everything. Whether that trade is worth it depends on how often you currently re-explain things and how often you currently redo work that was reported finished.

For a one-file change or a throwaway script, it is not worth it, and we would rather say so here than have you find out in week two.