Tutor:
You know what's on the other end now: a next-token guesser, trained on a mountain of text, shaped into an assistant, and blind to anything you don't tell it. That last part is the key to prompting.
Tutor:
Everything from lessons 6 to 8 reduces to one rule: the model's output is a function of the window, so prompt quality is context quality. Anthropic's framing is a brilliant new employee with none of your norms. Here is one request bare, then with the context that employee would need.
Bare request vs With who, why and lengthillustrative: made-up numbers that show the idea, not output from a real modelcompare
promptWrite a welcome message for my running club.
Welcome to the club! We're thrilled to have you join our community of passionate runners. Whether you're a seasoned marathoner or lacing up for the first time, you'll find support, motivation and friendship here. Together, we'll push our limits and celebrate every milestone…
Welcome! Nobody gets left behind here. Our Saturday slow group runs at chatting pace, walk breaks included, so come as you are. See you Saturday!
Tutor:
The right column came from adding one line: it's for nervous beginners, keep it friendly and under 40 words, and mention the Saturday slow group. Same model. Very different answer.
Tutor:
The habits, with the mechanism behind each:
Supply the facts. Names, dates, documents, decisions already made: if it's not in the window it doesn't exist (lesson 6). Anything recent or private must be pasted or retrieved (lesson 8).
Be clear and direct. Anthropic's test: hand the prompt to a colleague with minimal context; if they'd be confused, so will the model. Ambiguity is resolved by sampling, not by asking, unless you say otherwise.
Give the why. Motivation and audience let the model make the judgement calls you would. “For nervous beginners” did more above than any adjective.
Constrain the shape. Length, format, number of options. Unconstrained output follows the corpus's defaults, which are long and generic.
Use examples when style matters. Three to five, wrapped in <example> tags so they can't be mistaken for instructions. The model matches them closely, so the examples are the spec.
Structure and placement. In code, wrap instructions, context and inputs in XML tags; put the role in the system field; and with long documents (20k tokens and up), put the documents at the top and the question at the end, which Anthropic measured at up to 30% better on multidocument tasks.
Leave room for “I don't know” or a question. It's the cheapest hallucination guard there is. Then iterate: read, adjust, rerun. A prompt is an artefact you version, not an incantation.
Tutor:
Checkpoint time. Write a prompt for something you actually need: an email, a plan, an explanation. Hit all four checks, then run it.
Your turn
Run your prompt and the replies land here.
Tutor:
Finish this checkpoint and the course badge is yours. Two good next steps: Tokens & temperature goes deeper into tokens, odds and randomness, and Prompting that works turns these six habits into a toolkit.