Spending fewer tokens

What a token is

A token is a piece of a word, roughly four characters in English. Everything Claude reads in a conversation is counted in tokens: your messages, its answers, the files it opens. That count is limited. On a Claude plan, the more you read per session, the sooner you hit your usage limit; and an assistant that has read too much answers less precisely.

The goal is not to read less. It is to read only what serves the question.

The basic method rereads everything

In the method popularised by Karpathy, the assistant starts each session by reading the whole index: one line per note, with its summary. With 20 notes, it costs nothing. With 300, the index alone weighs thousands of tokens, paid every time, even to ask about one note.

What I measured on my own second brain

These figures come from my own second brain, about 90 pages, before and after each change.

Operation	Reread everything	With routing	Saving
Context loaded at session start: full index → theme map	≈ 4,400 tokens	≈ 1,650 tokens	−63%
Asking a question: index + 6 pages → 15 candidate names + 3 pages	≈ 17,000	≈ 6,000	−65%
Filing the day’s notes: whole wiki in the prompt → targeted retrieval	≈ 235,000	≈ 16,000	−93%
A question already answered	recomputed	filed answer reread	0 token to search

The four habits

1. A theme map instead of the full index. themes.md lists only note names, grouped by theme, with no summaries. Claude reads it, picks the theme and opens two or three notes. The kit is already set up this way.
2. Follow links instead of searching everywhere. That is what the graph from module 2 is for.
3. File the answers. An answer in qa/, still up to date, is reread rather than recomputed.
4. /clear when you change topic. Every message in a conversation is reread with the next one. A conversation that stacks five topics costs five times what it needs to.

Honestly: the biggest lever is the fourth

On a real working day, the architecture saves a few thousand tokens per session. A conversation where you change topic five times without /clear can cost over 200,000. In my own measurements, the architecture alone did not make up for that habit. Do both, but start with /clear.

Measure it yourself

Type /context in Claude Code: it shows what the conversation weighs. Ask a question with the kit as it is, note the figure, then /clear, and ask the same question telling Claude to read index.md in full first. The difference is the saving of habit 1, on your own notes.