PromptMe

Inspiration

AI systems are available in English and are getting better and faster over time. However, humans are just picking up on AI. Humans are the opposite of technology; we become slower as we age. There is a gap between the fast machine and the slow human. Not to mention, not everyone knows about the prompt engineer or has the time and resources to learn about it. As a second language learner with no engineering background, it took me a while to wrap my head around it. I read the article "Gliding, not Searching" by Professor James Intrilligator in my AI class. What if we have the same tool that helps us glide through any AI system to steer it to better results with just a few clicks?

What it does

Writing effective prompts for AI systems can be challenging. PromptMe helps users generate high-quality prompts with minimal effort. PromptMe analyzes initial prompts and provides targeted feedback to improve clarity, specificity, and formatting. It identifies areas where prompts could be shortened or restructured to better suit the user's needs. The tool then offers revised prompt suggestions that incorporate its recommendations. By simplifying and strengthening prompts, PromptMe helps users get the most out of their interactions with AI with less typing but more thinking.

How we built it

Utilized PartyRock playground to build and train PromptMe without code. There are four widgets:

User-prompt – get input from the user. Assistant-feedback (Claude) – evaluate user-prompt A better prompt (Jurassic-2 Ultra) – revise the prompt based on the assistant-feedback Try it out (Llama 2 Chat 70b) – provide answers to the newly revised prompt

All four widgets have different LLM models because of their function and suitability. First, I tried the preset LLM and prompt. The app gets trained with various topics from job applications, assignment questions, biography, etc. I tried different models for sections 2 to 4 to see the variation in the answers. I utilized the Retry button to provide LLM feedback on the result.

Challenges we ran into

Limited customization – there is no way to limit revised prompts to certain characters or words, especially when asking for creative prompts like creating images. The revised prompt gets so long that it exceeds the character limits of other AI systems.

Model limitation – the only way to train the model is by revising the user prompt and clicking the Retry button. The model doesn’t provide a consistent answer format for different subjects.

Prompt engineering – it is impossible to train the model to provide a better prompt based on the prompt engineering patterns.

LLM is not responding – there is an error when asking about “Design Justice,” which is unexpected.

Accomplishments that we're proud of

PromptMe closes the gap between speed, knowledge, and language. Users cannot recall all the words or ideas because they have limited memory. This is when PromptMe comes in. Users go through the feedback and revise prompts for insight. PromptMe empowers anyone with any background and skills to help interact with other AI systems so that they can focus on their main tasks.

What we learned

AI is not perfect. It still needs a human touch to make it perfect. PromptMe doesn’t always give the right results, but it does give users something to build on. It is all about gliding and not searching.

I have learned about different prompt patterns that users can apply when interacting with AI. However, these prompt patterns are subject to change as LLM becomes more sophisticated. PromptMe doesn’t stick to the patterning but rather gliding. This is important because users can learn about asking good questions without needing to learn prompt engineers.

You don’t need perfect English to train the model. I provide a lot of prompts with typo errors or grammatical errors.

What's next for PromptMe

Human-AI Interaction Explore various ways of human interaction with AI or human-AI collaboration for challenging tasks including tasks on ambiguous or controversial topics, AI in sensitive or regulated domains, etc. More importantly, advocating for design justice.

Feedback The only way that users can train or provide feedback to the model is through the Retry button. Explore ways to use various forms of human feedback that can best benefit model training.

Design

• Design and implement systems for collecting and integrating high-quality human feedback into AI systems.
• Incorporate design justice in the AI systems. I am surprised that “Design Justice” is not included in LLM because it is the design practice of intersectionality of the users. It should be incorporated during the design process to create an inclusive design for all.

Test and iteration Recruit participants with various backgrounds and abilities. This allows AI to learn how to provide better results.

Built With

• claude
• jurassic-2ultra
• llama2chat70b
• partyrock