Relax. We're not going to explain this chapter like a textbook — just like a friend talking you through it, in plain, easy-to-follow terms.
Let's think about it this way for a second
Remember three failure patterns. 1) Wrong text — an unrelated passage, someone else's notes. 2) Stale text — last year's syllabus, an old rule that's already been updated. 3) Too much text — 200 chat messages, 5 PDFs, a long tool log. The model can be confidently wrong. Drop a rotten fish into the soup and the whole pot smells like it. Polishing the prompt won't get rid of that smell.
Let's connect it to everyday life
A good turn — question + a short handbook excerpt + "say so if you can't find it." A broken turn — question + the entire past chat + a random wiki page + yesterday's wrong answer. In week one, don't let your context run past half a page. If you need more, search with RAG first, then add in a short excerpt. Don't dump a whole bag of salt into the soup.
Let's try it out together
Breaks when you pack:
- the wrong paragraph
- last year's rule
- the whole week of chat
Fix: keep 1–2 fresh, relevant notesTell apart the three ways context breaks down.5-Minute Try
Think of 4 wrong outputs. For each one, note whether it was wrong text, stale text, too much text, or a fuzzy prompt.
One Quick Heads-Up
A nice-sounding answer doesn't mean the context was right. Go back and re-read what was actually on the table before you trust it.