Assignment: The Example Library

Build two small programs that are taught entirely by example, one that is improved by pointing, and one that is repaired with a declaration. Your submission is the library of prompts, with notes on what each one taught and what it could not.

The rule from the walkthrough holds for the first two: no sentences in the pattern. The only instruction is the one line, sent first: “Complete the pattern. Output only the completion, with no explanation.”

Student instructions

1
Step 1: A classifier.
Choose a set of three to five labels for something you know well: messages to a club inbox, songs by mood, items to keep or throw out. Write four to six labeled examples and one incomplete one. Get the LLM producing the label and only the label. Then find the fewest examples that still work, and write one borderline case with the single example that teaches it.
2
Step 2: A procedure.
Write one worked example of how you work through a kind of problem, as a trace of thought, action, and result. Then give a new problem of the same kind. Check whether the LLM walks your steps, in your order, applying your tests.
3
Step 3: A refinement.
Take any prompt you have written for this course and generate with it. React by pointing: copy the sentences you liked, word for word, into a “preferred” list and the sentences you disliked into a “not preferred” list. Every sentence in both lists must be a direct quote from the response. Add both lists to the prompt and generate again.
4
Step 4: The declaration.
Find a case where one of your patterns misbehaves, or force one with an input far outside your examples. Repair it with the smallest declaration you can, such as a labels line or a single rule, and count the words you added.

What to turn in

For the classifier

The prompt you used, copied exactly as sent.

For the procedure

The prompt you used, copied exactly as sent.

For the refinement

The whole conversation: the original prompt, the first response, the prompt with the preferred and not-preferred lists, and the second response.

For the declaration

The prompt you used.
The response where the LLM got it wrong.
The updated prompt with the declaration added, and the number of words you added.

One paragraph. Where was the line between what examples alone could teach and what needed a declaration? State it as a rule of thumb you would give someone else.

We are looking at whether you understand what examples carry, not at whether every pattern worked the first time. A pattern that failed and was fixed with three words is a better submission than one that worked and taught you nothing.

Submitting prompts

Copy each prompt exactly as sent, including spacing. Copy each response exactly as received. Label them.