Useful tools · Watching the spend

Thirty-eight thousand tokens before anyone speaks

Our own configuration file is 1,897 lines and roughly 38,000 tokens, loaded into every session before a word is typed. The instinct is to cut it in half — don't. Every rule in it was bought with a real failure, and it rides in cached context that costs almost nothing to reload. The size is the point, not the problem.

There is a file that loads into every session you start, before you type a word. Ours is 1,897 lines. Yours is almost certainly longer than you think, and neither of us has read it end to end.

155,448 bytes, which is somewhere around 38,000 tokens, carried on every session and through every compaction. Two separate walkthroughs arrived at the same conclusion from opposite directions: that file is the largest single cost on a machine like this.

38,000
tokens · every turn
CLAUDE.md
155,448 bytes. Loaded whole, every session, survives compaction.
9
skills installed
Skills
Cost nothing until one triggers.
14
hooks configured
Hooks
Fire on events. No standing context weight.

One file outweighs the other twenty-three combined — it is the only one of the three that is prose loaded whole, rather than a trigger that costs nothing until it fires. Skill and hook counts measured 17 August 2026; both move most weeks.

The obvious fix is wrong

The instinct is to cut it in half. Do not.

Every rule in ours was bought with a real failure. The file says so about itself — a rule with a scar attached survives, "be careful" does not. The rule about never taking a host's word for its own health exists because a wedged session reported itself perfectly healthy while its work sat frozen. The rule about verifying a claim rather than remembering it exists because a roster count in that very file was wrong in both directions, twice.

Delete those and you do not have a leaner assistant. You have one that will rediscover each of those failures at full price.

What is actually wasteful is different

Three things, and none of them is a rule.

Duplication. The same instruction stated in three places, in three slightly different words, because three sessions each thought they were writing it down for the first time.

Reversals left in place. At least three passages in ours state a rule and then reverse it further down. Both halves are still there. A reader — human or otherwise — has to work out which one won, every time.

Procedure sitting in a memory file. This is the sharpest of the three. A memory file should hold knowledge: who you are, what is true, what was decided and why. Step-by-step procedure for a job you do occasionally is different — it belongs somewhere it loads when it is needed, not on every turn of every conversation for the rest of the year.

That split is the audit. Not "what can I delete", but "what is knowledge and what is procedure."

One honest limit

This is a measurement of a file, not of a bill. We have not proven the 38,000 tokens is our biggest line item — we believe it is, on arithmetic rather than observation, and we would rather say that than round it up into a finding.

The part that transfers to you

Open yours and check three things.

Does it say the same thing twice? Does it contradict itself, with both halves still standing? And is there a procedure in there — a sequence of steps for a job you do now and then — that could live anywhere else?

Those three passes cost you an hour and cost nothing to run. They are also the only kind of trimming that does not quietly remove the reason something works.

What to do

Read your configuration file top to bottom once, this week. Fix duplication and reversals, move procedure out, and leave every rule that has a scar attached. Then put a meter on it before you claim you saved anything.

Want this running in your own practice? Let's talk.