Prompt engineering advice

From https://www.youtube.com/watch?v=T9aRN5JkmL8

Based on the sources and our conversation history, the best methods for making what you want very clear to the language model revolve around clear and precise communication, much like explaining a task to a person who lacks specific context. The core idea is to "externalize your brain", meaning you must get everything you know and think about the task out of your head and into the prompt so that the model can understand it.

Here are specific methods discussed:

Write a Very Clear Description of the Task: This is fundamental. Think about it as giving a full set of instructions.
Strip Away Your Assumptions: Don't just write down what you already know. You must systematically identify the full set of information the model needs to understand the task and communicate it clearly. As one expert notes, many prompts are "conditioned on their prior understanding of a task," making them unclear to someone without that context.
Anticipate How Your Prompt Might Go Wrong: Unlike a human, models don't typically ask clarifying questions when instructions are ambiguous or inputs are unexpected. You need to think through potential edge cases and what a human might ask about.
Give the Model "Outs" for Unexpected Inputs: Since you've anticipated edge cases, provide explicit instructions for what the model should do when it encounters unusual or unclear inputs, such as outputting a specific tag like "unsure". This helps it handle inputs it doesn't know how to process according to your main instructions.
Leverage the Model to Critique Your Prompt: You can explicitly ask the model to tell you where your instructions are unclear or ambiguous before it attempts the task. If the model makes a mistake, you can show it what it got wrong and ask it to explain why it went wrong and even suggest edits to your instructions to prevent the mistake in the future.
Provide Full Context and Relevant Documents: Treat the model as smart and capable. Rather than "babying" it or dumbing down information, give it the full context or relevant documents it needs, like giving it a research paper if you want it to learn a technique described within it.
Describe the Actual Task and Context Precisely: Instead of relying solely on simple role-playing or metaphors (like saying "You are a teacher") that might be a shortcut to a similar task, clearly and prescriptively describe the exact situation you want the model to operate within. If it's embedded in a product or writing on behalf of a company, state that explicitly. Metaphors can sometimes help illustrate the type of analysis desired (like grading a chart as a high school assignment), but the focus should be on clearly communicating the task itself.
Use the "Temp Agency Person" Analogy: Imagine explaining the task to a competent person who just showed up from a temp agency with no prior context about your specific company or task. What would you need to tell them to get the job done right? This can help you identify missing information. Sometimes, simply speaking your instructions out loud and then transcribing that can result in a clearer prompt than you'd initially write.
Test Your Prompt with Another Person: A good way to check for clarity is to give your prompt to someone else who has no prior context about the task. If they can't understand what to do based on your prompt, it's likely unclear to the model as well.
Use Models to Help You Write Prompts: Experts increasingly use models to assist them in the prompt writing process. This can involve having the model interview you about the task to help you elicit and externalize the necessary information from your own brain, or generating examples to use in your prompt. This collaboration helps make your desired outcome "legible" to the model.