Your setup costs tokens every session

The part of your setup you pay for on every turn

When people tune Claude Code, they usually add: another instruction, another skill, another connected server. Each addition feels free, because you pay nothing at the moment you add it. The bill arrives later and keeps arriving, because anything that is always loaded gets read again on every turn.

That covers your CLAUDE.md, the listing entry of every installed skill, and the tool definitions from connected servers. The model reads all of it before it reads your actual request.

Three costs, not one

The first is tokens. Always-loaded material fills the context window before work begins, so you have less room for the task.

The second is usage limits. Anthropic's own usage-limit guidance lists conversation length as a factor, and a heavier baseline makes every conversation longer from the first message.

The third is the one people miss: attention. A model asked to follow two hundred lines of instructions follows each one less reliably than a model asked to follow twenty. Instructions compete. Outdated or contradictory ones make it worse, because the model has to guess which to obey.

Why skills are the fix

A skill is loaded on demand. Only its short listing entry stays in context every turn, and the full body arrives when the task calls for it. That changes what belongs where. CLAUDE.md should hold what applies to nearly every task: rules the code can't tell you, and how to run and test the project. Guidance that matters for one kind of task, such as how to write a release note or how to format a report, belongs in a skill.

A skill you never invoke is still not free. Its listing entry is in context on every turn. An unused skill is a subscription you forgot to cancel.

Audit it like a bill

You don't need to do this by hand. At the Claude Code prompt, /doctor runs a checkup. It reports findings first and asks before it changes anything. Among other things it looks at unused skills, servers and plugins compared with their context cost, slow hooks, and CLAUDE.md files with duplicates, content Claude could derive from the codebase, or always-loaded guidance it offers to move into skills. /doctor prompt-audit checks your instructions for outdated or conflicting content. /skill-doctor reports each skill's cost and usage, and flags skills that were never invoked.

Read what they find. Approve only what you understand. A cleanup that deletes a rule you needed is worse than the clutter.

What to do tomorrow

Update, open Claude Code in your project folder, and run the checkup:

claude update
/doctor
/skill-doctor

Then give your project instructions one rule to live by:

Keep here only what every task needs: rules that cannot be derived from the code, and the commands to run and test. Guidance needed only sometimes moves into a skill.

Check the result in a week. If nothing you removed was missed, it was cost without benefit.