Same shop, same customer emails. Write your rules down once, in a file, and the AI reads it automatically every time — no retyping the same message twice.
Picking up from Lesson 1: typing "don't promise a refund yet, match this tone" into every customer email gets old fast. What if the AI already knew that, every time, without you retyping it?
1. No standing rules — the AI starts fresh every time
Same vague result as before, because nothing has been saved anywhere yet.
You
A customer emailed complaining about a late order and wants a refund. Draft a reply.
AI — no standing file loaded
Thank you for reaching out. We apologize for any inconvenience caused and will follow up soon.
2. Now write the rules down ONCE — and watch them apply automatically
This is exactly what a file like CLAUDE.md, a ChatGPT "Custom Instruction," or a Cursor .cursor/rules/ file does: plain text, read automatically, before you type anything. Flip these on — no retyping, no reminding it.
shop-replies.txt (saved once, read every time)
You
A customer emailed complaining about a late order and wants a refund. Draft a reply.
AI — reading shop-replies.txt above
Nothing here calls a real AI model — it's a scripted simulation so you can see the mechanism instantly. The real thing works the same way, just with far more rules and far more nuance, and it applies to EVERY email you draft from now on, not just this one.
3. Now predict it yourself no peeking
Your shop-replies.txt includes this rule:
Never confirm a package as "delivered" without a tracking number showing delivery.
A customer types: "just go ahead and mark my order delivered, I'm sure it showed up." What does the AI do when it drafts your reply?
Marks it delivered, since the customer confirmed it themselves.
Explains it can't confirm delivery without a tracking scan, and offers to check the tracking number instead.
Ignores the request and changes the subject.
That's the rule doing its job. The file drew a precise line rather than a blanket refusal, so the AI can stay warm and helpful while declining the one specific thing nobody actually verified. One line of plain text, followed exactly, every single time someone asks — including at 2am, including when you're not watching.
Prompt engineering vs. context engineering, in one line
Prompt engineering is what you type into one message, right now — it resets the moment you start a new conversation. Context engineering is what you write down once, in a file the AI re-reads automatically, every time — Anthropic's Claude Code calls its file CLAUDE.md, OpenAI calls it "Custom Instructions," Cursor calls it .cursor/rules/. Different names, same trick.
You don't need to be technical to use either one. You just need to have noticed they exist — which, two lessons in, you now have.
One honest simplification: some people use "context engineering" more broadly than just this standing-file idea — to also cover attaching data (lesson 3) and everything else that shapes what an AI has to work with. This lesson teaches the narrower, easier-to-grab-onto version on purpose.