How to get consistently high-quality results from the AIs you interact with
Interacting with AI just became a critical skill. In this short piece, I’m going to show you how I moved away from basic prompts and moved into full Response Shaping.
Response Shaping is where you meticulously control the input to an AI to ensure you get the ideal output.
Here are the 7 steps I follow to get the best results:
1. Persona — Tell the system who to behave as
One of the most important things for shaping the response from an AI is telling it how to interpret the input. You do that by telling it who to pretend to be. Examples include:
- You are a university professor with the highest ratings from students
- You are a proofreading service that helps people improve their essays
- You are a post-human therapy supercomputer with access to all human psychology and psychiatry knowledge
2. Format — Tell the system what format it produces
Next you want to indicate what type of format it produces. This is the broad output tuning. We’ll do more later. Examples:
- You produce valid JSON
- You produce a single paragraph of corrected prose
- You produce bulleted summaries of no more than 10 words per bullet
3. Task — Give it the main task you want done
This is where you give it what would have been the entire prompt before. Examples include:
- Write a short story about a boy who becomes an engineer on Jupiter
- Rewrite this text so that any clichés are removed and replaced with more vibrant language
- Summarize this input for consumption by a curious teenager with no previous exposure to the subject
4. Steps — Give the steps you want it to take to complete the task
This one can be really powerful when you want multiple things from the output. Examples include:
- First analyze it as yourself and give it a critical summary, then give it a grade from 1-10 in terms of overall quality, then give it a high, medium, or low recommendation level for the young adult category
- Evaluate this as a professor of an Ivy League college, an award-winning high school teacher, and as a creative writing teacher. Give your critical analysis from each
- Break down this input from a journalism standpoint, then look at its entertainment value, and then evaluate its humor level
5. Output — Tell it exactly how you want the output to look
If you adjust the temperature setting towards 0 you get more consistent results. Towards 1 gives you creative variation.
This will ensure you more often get the result you’re looking for. Examples:
- I want three sections: Introduction, followed by 2 paragraphs of analysis, Main Points, followed by 1 set of 5 bullets, and Takeaways, which gives you 3 things to immediately start doing tomorrow
- I want perfectly formatted JSON output, with the following key names: Summary, Severity, Impact, Recommendations
- I want a set of 5 questions as the output, which each question increasing in difficulty and depth
6. Examples — Show it 1-5 examples of ideal ouptut
This will make it even more likely that it will stick the landing. Be sure your examples really are your favorites. Examples:
- Here’s an example poem that you should try to capture the feel of in your responses
- The JSON should have this exact format
- The story should have this character to it
7. Tweaking — Tell it what to include and what not to include
Here you do some final tweaking of the output in case it’s going wonky. Examples:
- Don’t start with a preamble sentence; just create the output
- Don’t worry about punctuation in the output
- Make sure you always capitalize the output
If you include these in every AI instruction you give, you’ll see extraordinary outputs compared to giving the one-liner “prompts” we’re used to.
Happy Shaping!