Pro Pro lesson
Prompting and checking answers
Real prompts, real answers: what helps a model, what doesn’t, and how to check what comes back.
This one’s part of How Models Learn Pro — every lesson, every interactive, and new ones as they land. About 12 minutes.
Already Pro? Sign in and you’ll go straight in.
Inside “Prompting and checking answers”
What you’ll learn
- Say what you want, and who it’s for
- Give it the context only you have
- Show examples and say what the output should look like
- Ask for the working, then check it
- Say what to do when it doesn’t know
- Check what comes back
Key terms
- Prompt
- The text you give a model: your instructions, any context or examples, and your question.
- Context
- Background facts in the prompt that the model couldn’t otherwise know, such as your situation or a document.
- Few-shot prompting
- Putting a few worked examples in the prompt so the model copies the pattern.
- Zero-shot prompting
- Asking a model to do a task with instructions only, without examples.
- Chain of thought
- Asking a model to write out intermediate steps before its final answer.
- Fabricated citation
- A reference a model invents, with a plausible title or author, that doesn’t exist.
- Sensitive data
- Personal, confidential or private information that shouldn’t be shared with outside services.
Sources
- Prompting best practices (Anthropic documentation) — Be clear and direct (the colleague test; be specific about output format), add context or motivation, use 3–5 relevant, diverse examples, structure prompts with XML tags.
- Reduce hallucinations (Anthropic documentation) — Explicitly giving the model permission to admit uncertainty “can drastically reduce false information”; verify claims with citations.
- Prompt engineering guide (OpenAI documentation) — Few-shot learning (include a handful of diverse input/output examples), include relevant context, use Markdown or XML to mark structure.
- Language Models are Few-Shot Learners (Brown et al., 2020) — GPT-3 (175B parameters) performs tasks from instructions and demonstrations given purely as text, with no gradient updates.
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Wei et al., 2022) — Chain-of-thought exemplars improved arithmetic, commonsense and symbolic reasoning, but only from ~100B parameters up; smaller models wrote fluent but illogical chains; no guarantee of correct reasoning paths.
- Large Language Models are Zero-Shot Reasoners (Kojima et al., 2022) — Adding “Let’s think step by step” raised a large model’s accuracy on MultiArith from 17.7% to 78.7% and on GSM8K from 10.4% to 40.7%.
- Language Models Don’t Always Say What They Think (Turpin et al., 2023) — Chain-of-thought explanations can systematically misrepresent the true reason for a model’s prediction.
- Fabrication and errors in the bibliographic citations generated by ChatGPT (Walters & Wilder, 2023) — 55% of GPT-3.5 citations and 18% of GPT-4 citations in short literature reviews were fabricated.
- ChatGPT and large language models: what’s the risk? (UK National Cyber Security Centre) — Queries are visible to the organisation providing the model; recommends not including sensitive information in queries to public LLMs.
- Qwen2.5-0.5B-Instruct model card (Qwen team) — The small instruct model (0.49B parameters, Apache-2.0) behind every answer in this lesson.
- Svalbard Global Seed Vault (Wikipedia, revision used for the claim checks) — Managed under a tripartite agreement among the Norwegian government, the Crop Trust and NordGen; stores duplicates of seeds from around the world.
- Decoding the ancient Greek astronomical calculator known as the Antikythera Mechanism (Freeth et al., Nature, 2006) — A real paper on the Antikythera mechanism, shown as an alternative to the model’s invented references.