enterpilot

Cutting Claude Code Token Consumption by 10-30%

GoModel audit log showing a Claude Code session, with the raw system prompt content visible in the interaction panel

Most people have no idea what their AI tools are actually sending.

I proxied Claude Code and read every request I was paying for.

Token consumption dropped 10-30% after four changes.

1. CLAUDE.md

It grows on its own. Nobody prunes it. It ships with every request.

Open it. Read every line. Ask one question: is this true for every task I run?

If not, delete it. Project-specific context belongs in the prompt, not in a file that rides along with everything.

2. The memory index

On macOS:

~/.claude/projects/-Users-[user]-projects-[project]/memory/MEMORY.md

Mine listed around 100 memories, each with a short description. All of it shipped on every request. None of it was relevant to the work I was doing.

MEMORY.md contents shown in the interaction panel, listing dozens of accumulated memory entries

Every one of these lines rides along on every single request, whether it’s relevant or not.

I deleted all of them.

3. Untracked files

The file list goes into the request. Build artifacts, scratch files, stray downloads - all of it costs money.

git status output in the interaction panel showing a long list of untracked scratch and results files

None of these tests/e2e/release-e2e-findings-*.md files were relevant to the task. They shipped anyway.

Clean the directory or add the noise to .gitignore.

4. Absolute paths

Claude Code sends full paths. Deep directory trees cost tokens on every tool call.

Mine is /Users/[me]/projects/[project]. Not optimal. Minor, but real.

The actual lesson

None of this was hard to fix. The hard part was seeing it at all.

Every one of these problems was invisible from inside the tool. They only showed up once I put GoModel - an open-source AI gateway I’m working on - in front of the API and read the raw traffic.

You cannot optimize what you cannot measure.