- 1What one message costsEvery message sends the whole window again; the fuller it is, the more each message costs.
- 2One subject, one windowA new subject gets a new window, not yesterday's 100,000 tokens carried into today.
- 3The stable part firstThe long part that does not change goes first and stays unchanged, so the cache hits.
- 4A fact on a page, not in the chatWhat must hold tomorrow goes in a file read at the start; a chat ends.
- 5Give it the sourcePoint it at the page, ask for the sentence, and let it say the page does not say.
- 6The smallest model that does the jobModel and effort chosen per task; thinking tokens are paid tokens.
- 7Read lessSearch, then read the lines found; everything read stays in the window.
- 8Check what it saysSums by a tool, a claim against its line, and a correction where it lasts.
Use any AI well, in eight short habits.
People pay for tokens and get worse answers from habits nobody taught them. Each lesson is one habit, the reason under it, and its arithmetic at Anthropic's list prices. Each says how to keep the habit with instibrain, and without it.
The prices are Anthropic's list prices for its API. On a Claude plan you see no bill, and the same habits make your usage limits last longer.
Eight habits, from what one message costs to checking the answer.
Lessons 1 to 3 are about the window and the bill: what a message reads, when to start again, and how the cache works. Lessons 4 to 8 are about better answers: where a fact lives, what to give the model, which model, how much to read, and how to check.
The eight habits, and the number under each.
Prices are for Claude Opus 5.5 on Anthropic's API, per million tokens: $4 to read, $5 to write the cache, $0.20 to read from it, $20 to write an answer.
| Habit | The number | Lesson |
|---|---|---|
| Watch how full the window is | A message in a 100,000-token window reads $0.40 of tokens; one in a 6,000-token window, $0.024 | 1 |
| New subject, new window | 100,000 old tokens carried through 20 messages: $8.00, or $0.88 from the cache; a new window: $0 | 2 |
| Keep the stable part first and unchanged | A cache read is a twentieth of the price on Opus 5.5 and Sonnet 5.5, a tenth on most other models; it lives 5 minutes, or an hour at twice the write price | 3 |
| A fact goes on a page, not in the chat | A rule typed in the chat is gone at the next window | 4 |
| Give it the source | Ask for the sentence, and let it say "not in the file" | 5 |
| The smallest model that does the job | Haiku 5.5 costs 40 times less than Opus 5.5 for prompts up to 100,000 tokens, 8 times less above; thinking bills as output | 6 |
| Read less | A 5,000-line log is about 60,000 tokens, 6% of the window; the 20 lines a search finds, about 240 | 7 |
| Check what it says | Which file, which line: an answer you act on names both | 8 |
Your brain on your computer: spaces, specialists, tools, the button, the inbox, and where a lesson goes.
with a brainBuilt for AIA brain loads in three layers: who your AI is, what exists, then one page at a time.
every seriesGuidesThe four kinds, what you need, and how a lesson works.