What You Will Be Able To Do

• Write prompts with role, task, context, format
• Use examples to show a pattern
• Structure long prompts with labels
• Improve prompts through small tests
• Recognize and fix common failures

Think Of A Capable New Colleague

• Reads fast and writes well
• Knows nothing about your situation
• Cannot see your goals or audience
• Weak output usually means missing information

Roles And Messages

• System message: tone, scope, limits
• Chat tools hide it; APIs expose it (week 10)

system:    standing rules and role
user:      the current request
assistant: the model's reply

A Role That Helps

• Role sets audience, style, and depth
• A role adds no new knowledge

Weak:   You are an assistant.
Better: You are a patient tutor who explains
        concepts to adult beginners using
        everyday examples. Answer only from
        the supplied text.

Four Parts Of A Strong Prompt

• Task: a direct instruction
• Context: audience, purpose, facts
• Format: list, table, JSON, length
• Constraints: tone, scope, limits

Say What You Want

• Positive targets are easier to follow
• Not: do not use jargon
• Better: write for a high school student
• Make it checkable: 3 bullets, under 20 words

Show, Do Not Only Tell

• Zero-shot: none. One-shot: one. Few-shot: several

Classify the review as positive or negative.

Review: The battery lasts all day.
Label: positive

Review: It broke after two days.
Label: negative

Review: Fast delivery but the screen is dim.
Label:

Reasoning Step By Step

• Chain-of-thought helps on multi-step tasks
• Some models reason on their own; test it

A shop sells pens at 3 for 2 dollars.
How much do 12 pens cost?
Think through the steps, then give the
answer on a last line starting with ANSWER:

Structure Long Prompts

• Labels or tags separate the parts
• Material first, question after it

<role> ... </role>
<rules> ... </rules>
<example> ... </example>
<document>
  ...pasted text...
</document>
Question: ...
Reply in the format: ...

Asking For Structured Output

• Name the fields and allowed types
• Say nothing else should be returned
• Validate the result in code

Return only JSON with these keys:
{"title": string, "actions": [string],
 "deadline": string or null}

Iterate Like An Engineer

• Run on several real inputs
• Find the pattern in the errors
• Change one thing at a time
• Rerun the whole set and compare

Failure Patterns And Fixes

• Vague output: add audience, length, format
• Invented facts: supply the source, answer only from it
• Ignored rule: shorten, state once, place near end
• Wrong format: show an exact sample
• Too agreeable: ask for flaws and counterarguments

Grounding The Answer

• Supply the facts yourself
• Allow the model to say it does not know
• Ask for the supporting quote

Answer using only the text in <document>.
If the answer is not in the text, reply:
"Not stated in the document."
Quote the sentence you used.

Where Prompting Stops

• No wording can supply facts the model never saw
• Large document sets exceed the context window
• Some tasks need actions, not only text
• Retrieval and tools fill these gaps

Recap

• Be clear: task, context, format, constraints
• Show examples and structure long prompts
• Iterate on a small test set
• Know the failure patterns and their fixes

Next Week: RAG And Tools

• Retrieval-augmented generation: find the right text first
• Embeddings and search from week 4 put to work
• Preview of tools that let a model take actions

Questions And Exercise

• Exercise: Improve a prompt on a test set
• Share one result at the next session
• Bring one question to week 7