Practice

What students will do

Students will teach an LLM a task using nothing but examples, then teach each other the same way and find out which is harder. The point is to feel how much a handful of examples can carry, and to notice the moment where an example does what no rule could.

Notes for teachers

Use the Pattern Trainer at /exercises/pattern-trainer or a chat tool with the single instruction sent first: “Complete the pattern. Output only the completion, with no explanation.” Enforce the rule: no sentences in the pattern. Students will want to explain. Do not let them. The discipline is where the learning is.

Part 1 · The secret rule

Step 1 · Teach a rule without stating it

1
Each pair invents a secret rule for sorting words, numbers, or short phrases into two groups.
Words with double letters versus without. Numbers that are one more than a square. Phrases you could say to a dog. Keep it findable but not obvious.
2
Write six labeled examples, three per group, then one left incomplete.
No sentences. No hint of the rule.
3
Send it.
Did the LLM put the incomplete example in the right group? Send two more incomplete examples to be sure it found the rule and not a coincidence.

Discuss

If the LLM got it, which examples did the work? If it did not, which example would you add?
Swap patterns with another pair, without the rule. Can they figure out the rule from the examples alone? Who found it faster, the class or the LLM?

Part 2 · The boundary

Step 2 · Find the edge and teach it

1
Take your pattern from Part 1 and write the most borderline case you can.
Something that could reasonably go either way.
2
Send it and record which way the LLM went.
3
Add exactly one example that settles the boundary, and send the borderline case again.
Do not add a rule. If one example is not enough, add a second, and note that it took two.

Discuss

Say the boundary out loud as a rule. How many words did it take? How many did the example take?

Part 3 · The trace

Step 3 · Teach a way of working

1
Choose a kind of problem you actually solve.
Getting a group project unstuck. Figuring out why a recipe failed. Deciding which of three offers to take.
2
Write one worked example as a trace: problem, thought, action, result, thought, action, result, answer.
Then a new problem, ending at the first thought line.
3
Send it and read the trace the LLM produces.
Does it think the way you did? Does it take the actions you would take?

Wrap up

Where did examples do something a rule could not? Point to the exact example.
Where did examples fail, and what was the smallest thing that fixed them?