A 2023 Guide to Prompting ChatGPT: Instructions, Examples, and Limitations

In this February 2023 tutorial, Matthew Berman introduces ChatGPT as a conversational interface to a language model. He describes applications such as answering questions, drafting, and translation. A fluent answer can still be incorrect or misleading, so important facts and generated outputs need checking.

Berman defines prompt engineering as writing clear instructions for the response one wants. His simple example begins with 'the ocean is,' then narrows the response to a single word and finally a color. Adding constraints can make an answer more useful, although it does not guarantee correctness.

He breaks prompts into four possible parts: an instruction, context, input data, and an output indicator. For example, a user could supply a document, ask for a summary, explain the intended audience, and request a short list. Not every task needs all four elements, and formatting should help distinguish instructions from supplied material.

For summarization, Berman asks for an explanation of a microwave and then a shorter version. He notes that requesting a one-sentence explanation up front can avoid an extra step. The desired length, audience, and form are useful constraints; any technical explanation should still be checked against reliable sources.

Information extraction asks a model to identify details in supplied text, while question answering can be shaped by specifying an audience or response length. Berman also demonstrates classifying movie reviews as positive, negative, or neutral. Ambiguous wording and missing context can make these tasks error-prone.

Berman describes conversational roles and customer support, then discusses using GitHub Copilot to generate and explain code. He says the tool saved him time, and shows a simple logical reasoning example. Generated code and explanations still need testing and review, especially when used in real systems.

Zero-shot prompting gives an instruction without examples; one-shot gives one example; few-shot supplies several. Examples can communicate a pattern that is difficult to specify in a short instruction. They should be representative and should not contain sensitive information unless it is appropriate to share.

The video introduces chain-of-thought prompting through worked examples and asking for an explanation. It then describes self-consistency: sample different reasoning paths and choose an answer by consistency across them. The approaches can help in some settings, but multiple responses can share the same error, and explanations are not proof that an answer is correct.

For a word problem, Berman says an early response incorrectly answered that a 70-year-old's sister was 73, then a later attempt using self-consistency answered 67. The age difference is three years, so 67 is correct. He acknowledges that he did not record the initial failure and could not reproduce it afterward; this anecdote is not a controlled evaluation.

Generated knowledge prompting asks the model to produce potentially useful information before using it in a subsequent answer. That can help organize a response, but invented or incorrect intermediate material can also contaminate the result. Checking against independent sources remains important.

The final section covers adversarial prompting. In Berman's customer-support example, an attacker tries to make a model contradict a store's return policy; other examples aim to reveal an application's hidden instructions or redirect its goal. These examples illustrate prompt-injection risks when untrusted content is treated as an instruction, not reliable recipes for any current model.

Berman also describes attempts to bypass model guardrails through framing a request differently. Defenders should not assume a prompt alone protects an application: access controls, validation, and careful handling of untrusted content matter. He points viewers toward additional learning resources, while a short tutorial alone cannot confer expertise.