Over the past couple of years I benefited a lot from the experience of others in the world of AI.
Today it’s my turn to return the favor and share some of the magic words I use to get better results and stay ahead of the curve. The first advice is to go back to the old ways of working and ask your agent to RTFM.
Use published research to your advantage
Humans in academia are very curious creatures. They love to read, and they love to write. And they love to share their work with the world. This is amazing for us, because it means that we can ask Codex to find, read, and apply the published research to our problem. This is especially true for problems where what’s good is obvious when we see it, but hard to define in words.
Problems like:
- Good design that “feels right”
- Good code comments
- Copywriting that converts well
- Good UX that is intuitive and easy to use
- A social media photo that goes viral
… … … the list goes on and on.
The magic words are:
Find and read the published research on {problem}. Then, apply what you learned to {task}.
I often use it when I get stuck on something that feels wrong, but I can’t explain why with words. The solution is far from perfect, but it will give you a new direction to explore and a cool list of research papers to skim through at the end of your day.
Embrace ImageGen
It started with selfies, then the AI slop posters, but for those of you paying attention you probably noticed that it can do a lot more. It can redesign UI elements, generate new icons, create visual assets… … …
The most common use cases for me are:
- ImageGen + the research prompt from above.
This allows me to explore different directions for a given problem. - Create different designs for a given UI element.
Codex can take a screenshot of your UI and use ImageGen to create different versions for a given UI element without redesigning the whole screen. And ofc, the new designs will fit with the rest of the UI elements. - All types of social media content.
From slideshows, to video editing and everything else you can think of.
As long as you deliver a value, people are willing to look past the signs of AI generated content.
There’s no magic words here, it’s up to you to explore the different possibilities and find the right combination of prompts.
Specify the writing style
Everything you publish online needs to have a goal and a writing style. The goal could be something like knowledge sharing, brand awareness, lead generation, sales, etc. The writing style controls the tone, voice and personality of your content.
If you’re writing for a technical audience, you might want to try writing styles such as ASD-STE100 Simplified Technical English and google developer documentation style guide .
The magic words are:
Write {content} in the style of {writing style}.
For everyday communication The Chicago Manual of Style is a good starting point.
For creative writing you can use prompts like: What would {author} say about {topic}?.
The author could be a famous writer, a fictional character, or a generic persona like “professional copywriter”, “your uncle”, “karen from HR”, etc.
Finally, if all of this sounds like a lot of work, you can take the lazy route and say something like: Give me 10 options for {content}, each of them with a unique and distinct voice. These tend to produce the worst results, but I find them helpful when I’m in brainstorming mode.
Do less, complain better
Speaking of lazy mode… … …
AI is lazy, just like you. A lot of the time it will make assumptions and present you with a solution that is built on top of them. One way to fight back and get better results is to write a short list of:
- what you liked
- what you didn’t like
- what you expected to happen
The agent will review and hopefully, deliver the desired result.
If that sounds like too much work, you can also try writing a short review of the solution and complete the paragraph with the magic words: Would you like to challenge yourself to do better?
Most of the time, the agent will respond with a better solution. On rare occasions, it will build several different versions, judge them and decide that they’re not an improvement over the current solution.
Use skills
The best outcomes from AI are achieved when we give precise instructions using the vocabulary of the domain we’re working in. Usually, you would pick up a book, read a few articles or ask a domain expert to help you with the task. These days you can use a skill to skip a lot of the learning curve and jump right into the result.
If you’re in the tech world you might be aware of Emil Kowalski’s design skills and the skills from people like Adam Lyttle. These are great examples of how hard learned knowledge can be distilled into a skill and shared with the world. Use them as a starting point and experiment until you create something that works for your project.
One more tip: sometimes Codex can read and apply skills without installation. Just ask it to read the skill and apply it to your task.
You can combine this with the web search and let codex automatically find and apply several skills related to your task.
In the end, all you have to do is to look at results and pick a winner.
It’s okay to give up
Sometimes even the greatest magic in the world can’t deliver the results you wanted. If you find yourself repeating instructions and seeing similar results over and over again, it’s time to stop.
Step back, take a break, drink some coffee. Try again. or don’t.
Nobody said AI has to be your only item in your toolbox. It’s okay to give up and move on to something else.