Read less.
Point your AI at what it needs, not at everything. Every file it reads and every line a command prints stays in the window for the rest of the session.
new words
- Tool result
- What comes back when the AI runs a command or reads a file. It goes into the window like a message.
- Search first
- Find the lines that matter, then read only those.
- Progressive disclosure
- How a brain loads into an AI: who it is first, then an index of what exists, then one page at a time.
the lesson
Ask your AI to "look at the log", and it may read all of it. A log line of about 9 words is about 12 tokens, since 1,000 tokens hold about 750 words:
5,000 lines × 12 tokens = 60,000 tokens = 6% of a 1,000,000-token window
That is Ctx 6% from one command. Read once at $4 per million, it costs $0.24 on Opus 5.5. It then stays in the window, and every later message reads it again, at $0.012 each from the cache. It also adds to the context rot of lesson 2.
Ask instead for a search: the 20 lines that name the error are about 240 tokens, a 250th of the log.
| Instead of | Ask for |
|---|---|
| "Look at the logs" | "Search the log for the error, and read only those lines" |
| "Read the whole folder" | "Find the file that sets the due days, then read that file" |
| Pasting a long document | Naming the file and the section |
With instibrain, your AI reads this way by design: who it is first, then a map of what exists, then one page at a time as the task needs it, progressive disclosure. Without it, ask for the search yourself.
try it
In a folder with some files, type this and press Enter:
Find where this folder mentions invoices. Search first and read only the lines you need. Then tell me how many lines you read.
you should see
- A search first, then a few lines read, not whole files.
- An answer with the file and the line, and a small count of lines read.