<ol><li>Check if conditions are fulfilled: Make sure your prompt contains what should be done if GPT is unable to do the task. An example would be "Respond with \"no steps provided"\ if there are no bullet points"</li><li>Few shot prompting: Giving the model a few example of what you want, a conversational style, or an explanation in the style of a professor goes a long way in making it give content you want.</li></ol>A fundamental principle of CS applies here, GIGO. Garbage In, Garbage Out. If we give it the right info, it'll give us the right response.

Instruction tactics

<ol><li>Specify the steps: By specifying the exact steps needed, we get a much clearer output. An example would be to summarize a research paper, convert to french, list each heading out and convert it to a JSON object.</li><li>Instruct the model to figure out its own solution: Giving the model steps to do to get the output, and defining how you want the solution to be will give you your required output. An example would be to ask it solve a math problem by itself, and then compare it to a solution delimited by `` and check if it is true or not.</li></ol>

Thinking tactics

ChatGPT is all the rage, and with that comes a plethora of Prompt engineering cheatsheets that are supposed to improve how you work with it.Let's figure out how these prompts are built, instead of just directly copying them This is a series of posts, each covering specific concepts

Fundamental Guidelines for prompt engineering with ChatGPT

ChatGPT doesn't think like a human, it's a computer program after all. In order to have it respond correctly, we first define the base conditions on which it will act. A sandbox of sorts.There are four basic tactics to apply:<ol><li>Use delimiters: Tell GPT to "Summarise the text between ```" or any other delimiter of your choice. This prevents anyone from engineering on your own prompts. </li><li>Ask for structured output: Most times we need structured data, as a CSV, JSON or even an HTML file. By asking GPT to reply in a certain format, we get the data that we want</li></ol>(Continued in the idea below)

Clear and Specific instructions - Principle 1

GPT doesn't think like a human, but it does need to think in a colloquial sense. For the best responses, we need to give it time to reason and think.By asking it to validate or sample an existing statement, it will understand and analyse it based on the POV that was presented in the text. But if you let the model understand it based on it's own terms and then ask it to infer, it'll give you a different POV altogether.We can do this by applying the following tactics

Let GPT think - Principle 2

ChatGPT is a powerful tool, and understanding how to ask it is knowledge that we all should have. "Give a man a fish, and hell be sated for a day. Teach him how to, and he'll never be hungry in his life". The same philosophy applies here.

Instruction tactics

Similar ideas to Instruction tactics

7 Tips for Saying No Effectively