Context and tokens: spend less and get better answers

What you'll see in this lesson

1. The cost isn't what you ask, it's what you carry

   Every new message travels with the whole conversation that came before it, and you pay for that whole package.

2. The model doesn't remember: it rereads everything

   On every turn, the app sends it the full conversation again, starting from the first message.

3. The same task, three and a half times cheaper

   Twenty turns in a single session versus four sessions of five turns.

4. One task, one session

   When you change subject, start a clean session.

5. Leave a note for the next session

   A short file with the status of the work gives you continuity without dragging the whole history along.

6. Short instructions: you pay for them every turn

   CLAUDE.md and AGENTS.md, the instruction files the agent reads at the start of each session, travel with every message.

7. Search before opening

   Don't let the agent read half the project to find one function.

8. Reasoning: medium by default

   Thinking harder costs more tokens and more time. Save it for the hard stuff.

9. Filter the noise from tests and builds

   Tests and builds, commands that check the code, can spit out thousands of lines that go into context in full.

10. Six habits that lower the bill

    None of them needs a new tool, only discipline.