Most of what looks like the model being unhelpful is really the model working from the wrong material. It has no background knowledge of your company, no access to the file you didn’t mention, and no memory of yesterday’s conversation. It has exactly what is in the window in front of it, and that window has a hard edge measured in tokens.
The failure this produces is the recognisable one: a session that starts sharp and gets vaguer. Halfway through a long piece of work, Claude begins asking about something you settled an hour ago, or quietly contradicts a decision you both made. Nothing broke. The desk filled up, and the earliest things slid off the far side of it. The second, less obvious failure is clutter — a window stuffed with a previous unrelated task, where the important detail is technically present but drowned.
Both fixes are cheap, and neither is clever wording. Give it the material explicitly rather than describing it: drag the files into the chat, or point it at another folder it can’t see yet. Then keep sessions to one job — a fresh chat between unrelated tasks costs you nothing and prevents most of the drift. That deliberate choosing of what goes in and what stays out is what people mean by context engineering, and it is the highest-leverage habit on this site.
One distinction is worth getting right early, because it causes real confusion: the context window is short-term and the session ends it, while memory is what persists across sessions. “Claude forgot” and “Claude never knew” are different problems with different fixes, and one of the guides below is entirely about telling them apart. Another draws the line between prompting and context — a useful frame, since asking well only helps once the right material is in the room.
The words
- Context window
- How much the model can hold in mind at once — your question, the files it's looking at, and everything said so far — measured in
tokens. Think of it as the model's desk space. When it fills up, the oldest things start to slide off the edge. - Context engineering
- The discipline of deciding exactly what goes into the model's limited context window — which files, which past messages, which instructions, which looked-up documents — so it has everything it needs and nothing that distracts it. Think of it as packing a small bag for a trip: the right items matter more than the most items, because space is finite and clutter hurts. It's the 2026 evolution of prompt engineering, focused on the whole working set rather than a single clever sentence.
- Token
- A chunk of text — very roughly three-quarters of a word. AI models don't read letters or whole words; they read tokens. The word
unbelievablemight be three tokens:un,believ,able.
How it works in Claude
Commonly confused with
Two things people mix up here, and the difference that actually matters.
Read more
How to do it
| What you want | On Desktop (no code) | Command line (Power Track) |
|---|---|---|
| Give Claude more context | Drag extra files into the chat | /add-dir ../assets |
-
/add-dirGive Claude access to another folder beyond the one you started in.