Told every session

About this video

On Monday you tell your agent where the reports are, how to write numbers, and which folder it must never touch, and it does a good job. On Tuesday you start a new session, and it searches the whole computer for the reports, writes 12500 without a comma, and offers to tidy up the archive. Nothing broke. It simply never knew.

A language model learns only during training; after that its numbers stay the same, and between requests it keeps nothing. So in any session it knows only what is in that session's context: the app's own instructions and your new message. Monday's conversation isn't there.

An instruction file is a plain text file of notes about a project that the agent's program, the harness, reads at the start of every session and puts into the context, before your first message. In Claude Code it is called CLAUDE.md, and it can live in your home folder (your own notes, for every project), in the project's folder (shared with the team) or in a subfolder (loaded the first time the agent works there). Many other agents read AGENTS.md, an open format for the same idea, and Claude Code reads it when a project has no CLAUDE.md.

Every line is sent in every request of every session. That costs tokens, leaves less room for the task, and competes for attention: ten sharp lines are followed more reliably than three hundred loose ones, and Claude Code's documentation suggests under about 200 lines a file. So write lines you could check, like the exact command to run or the folder never to change, not "write good code". And remember that it's guidance, not a lock: the model usually follows it. This video explains the instruction file on Knaph.

Transcript

1. 0:00Monday. You tell your agent where the reports are, how to write numbers, and what never to touch.
2. 0:07It does a good job.
3. 0:10Every answer follows what you said.
4. 0:18Tuesday. You start a new session.
5. 0:22It searches the whole computer for the reports, writes the number without a comma, and offers to tidy up the archive.
6. 0:30Nothing went wrong. It simply never knew.
7. 0:36Why? A model learns only in training. After that, its numbers don't change.
8. 0:42So on Tuesday it knows only what's in Tuesday's context: the app's instructions, and your new message.
9. 0:49Monday's conversation isn't there.
10. 0:52To know something every session, it has to be put into the context every session.
11. 0:58So write it once, in a file. An instruction file.
12. 1:03At the start of every session, the harness reads it and puts it into the context, before your first message.
13. 1:11Monday, Tuesday, Wednesday: every session begins with the same notes.
14. 1:16The model doesn't learn your project. It's told about it, fresh, each time.
15. 1:22In Claude Code the file is called CLAUDE.md.
16. 1:26Your home folder: your own notes, in every project.
17. 1:30The project's folder: its notes, shared with the team.
18. 1:34A subfolder: loaded the first time the agent works there.
19. 1:39Many other agents read AGENTS.md, an open format for the same idea.
20. 1:44But every line is sent in every request of every session.
21. 1:49A long file costs tokens each time, and leaves less room for the task.
22. 1:55It competes for attention too. Ten sharp lines are followed better than three hundred loose ones.
23. 2:01Claude Code's documentation suggests keeping each file under about two hundred lines.
24. 2:08So what goes in? Lines you could check.
25. 2:12Not "write good code", but the exact command to run after editing a report.
26. 2:18Not "be careful with files", but which folder never to change.
27. 2:23And remember: it's guidance, not a lock. The model usually follows it.
28. 2:28It doesn't learn your project. It's told, every session.