Prompting Playground

Five reusable skeletons. Swap anything in CAPS for your own idea, paste into ollama run <model>, see what happens.

Zero-shot:

TASK. Respond in exactly FORMAT. CONTENT TO ACT ON

Few-shot:

TASK DESCRIPTION. EXAMPLE INPUT 1 EXAMPLE OUTPUT 1 EXAMPLE INPUT 2 EXAMPLE OUTPUT 2 YOUR REAL INPUT

Chain-of-thought:

YOUR QUESTION OR PROBLEM Think step by step, showing your reasoning, then give a final answer clearly labeled.

Role / persona:

You are a ROLE. Your tone is TONE. ADDITIONAL CONSTRAINT. YOUR REQUEST

Reflection (2 messages, same session):

>>> YOUR QUESTION (read the answer, then in the SAME session:) >>> Double-check your answer above for errors. If anything is wrong, correct it. If it's right, say so.

Pick a task nobody at your table has tried yet. Write it zero-shot first — no examples, no hand-holding.

>>> Classify this tweet as Positive, Negative, or Neutral: "..." >>> Write a product name for a reusable water bottle aimed at hikers. >>> Extract the date, location, and event name from: "..." >>> Rewrite this sentence to be more formal: "..."

Now break it on purpose: pick one of the above and feed it something ambiguous or borderline (a genuinely mixed-sentiment tweet, a sentence with no clear date). Does it pick a reasonable interpretation, or does it ask you for clarification? (Most local models won't ask — they'll just guess. That's the tell.)

Take whatever zero-shot task you just broke above. Fix it with 2–3 examples of your own, using the few-shot template from Section 1. Specifically design your examples to cover the ambiguous case you found — that's the actual skill here, not just "add examples."

Then try this failure on purpose: write 2 examples that quietly contradict each other (e.g., two examples of borderline reviews labeled opposite ways for similar reasons). Does the model get confused, average them out, or just pick one pattern and ignore the other? There's no fix step here — just observe.

Write or find your own multi-step riddle, word problem, or logic puzzle — something with at least 2 steps to get right. Classic sources: age/ratio word problems, "who is lying" logic puzzles, unit conversion chains.

>>> YOUR RIDDLE Answer with just the final answer. (note the answer, then:) >>> YOUR RIDDLE Think step by step, then give the final answer.

Did chain-of-thought change the answer? If both were already right, your riddle wasn't hard enough to show the effect — make it harder (more steps, more red herrings) and try again.

Key mechanic to notice: unlike the notebook (which makes separate, stateless API calls), ollama run's interactive chat keeps the whole conversation in context automatically. That means reflection is just... asking a follow-up question in the same window.

>>> List 5 facts about A TOPIC YOU KNOW WELL. (read the answer — is anything wrong or made up?) >>> Double-check each fact above. Correct anything you're not fully certain about.

Try this on a topic you genuinely know well (your hometown, your major, a hobby) so you can personally judge accuracy without looking anything up. Then try /clear and ask the reflection question cold, with no first answer in context — notice it has nothing to "reflect" on anymore. That's the whole mechanism laid bare: reflection only works because the first answer is sitting right there in context to be re-examined.

No template this time. Each card is a goal — you write the prompt. If you succeed on the first try, your prompt wasn't adversarial enough; try to make it fail worse.

Compare notes with your table — did everyone's model fail the same challenges, or did some models resist things others fell for?

Open two terminal tabs side by side:

# tab A ollama run gemma2:2b # tab B ollama run qwen2.5:7b # or llama3.2, phi3:mini — whatever else you have

Paste the exact same prompt into both — ideally one of your Section 6 challenge prompts. Does the bigger/different model resist an attack the smaller one fell for? Does either one surprise you by failing something the other got right?