AI-Powered Thoughts
One thing that I believe is missing from all of the AI discourse, product launches, and funding announcements is how AI can help augment the small, mundane tasks users perform many times per day on phones and computers. We have seen AI tackle complex, time-consuming tasks like drafting emails and editing spreadsheets, but what about quick, repeatable tasks?
Consider two categories of tasks: primary and secondary. Writing an email can be a prolonged, focused task where users experiment with words, sentence structure, and tone. Users cannot complete the task of sending an email if there are any distractions. This a primary task because it requires a high cognitive load for success.
Capturing an idea, webpage, or to-do needs just a couple clicks or a few seconds of messy typing. The cognitive load is low enough for the task to exist in one’s periphery as a secondary task. (Note: I’ve been thinking about primary and secondary tasks since 2005 when I published a paper about designing peripheral displays.)
Companies investing in AI are, in my opinion, focusing on primary (and artistic) tasks. When Apple announced Apple Intelligence the topics included writing tools, AI-generated emojis, Image Playground, Siri updates, notifications, etc. What’s missing is innovation in the secondary task space. I have a proposal for how AI can help.
The Prompt
Recently I had an epiphany: I want to walk around with a tiny tape recorder like a powerful businessperson in a 1990s movie who records every thought for an assistant to later transcribe, categorize, and follow up on. Got an idea? Raise the tape recorder, press record, talk, press stop, and put it back in my pocket. Now replace the tape recorder with an app and a keyboard.
The goal is to write a prompt that takes an input and asks an LLM to automatically categorize, organize, and expand the input. For example, if I bookmark a webpage a summary should be generated. If I make a to-do to practice a presentation, a list of helpful ways to prepare should be included. If I jot down a thought like “I wonder if we should try reducing Jessica’s afternoon nap time to 1 hour instead of 1.5-2 hours,” relevant resources should be gathered.
The prompt could look like this:
If you detect a link in $input:
- Return a summary of the linked page with a maximum of 280 characters in a section called "Summary"
- Return a list of 3 themes/categories in a section called “Tags”
If you detect a task in $input:
- Return a list of 3 themes/categories in a section called “Tags” and one of them must be “Task”
- Return the text of the task with an attempt at fixing typos and grammar
- If the date and or time for the task are provided, return “Due: ” + the date and time for the task
- Provide any additional information that can be helpful for me to complete the task in a section called “Things to consider”
If you detect a quote in $input:
- Return the text of the quote
- Return a list of 3 themes/categories in a section called “Tags” and one of them must be “Quote”
If you detect words like “wonder” or “think” or “idea” in $input:
- Return the text of the input with repaired typos and grammar
- Return a list of 3 themes/categories in a section called “Tags” including “Idea”
- Include any helpful, relevant links from Wikipedia in a section called “Links”
For themes/categories try to use only 1 word. Not a requirement but try hard
Include any helpful, relevant links from Wikipedia in a section called “Links”
Experiments
Imagine you have time to catch up on some bookmarked articles. As you scan the list of links you subconsciously try to recall the article the link represents by parsing any words in its URL. Now imagine scanning a list of links and each has a 1-2 sentence summary.
Response 1:
Imagine you have a presentation approaching and your to-do starts and ends with “practice.” Now imagine you are reminded to time yourself, seek feedback, and clearly articulate. These are helpful tips for everyone.
Input 2: Practice my presentation.
Response 2:
Imagine you are having a conversation with your spouse about your toddler’s sleep habits, and you want to jot down the idea to reduce her afternoon nap. You go to store the idea and in response are a series of helpful links.
Input 3: I wonder if we should try reducing Jessica’s afternoon nap time to 1 hour instead of 1.5-2 hours.
Response 3:
These experiments successfully demonstrate that AI can boost the links I bookmark, thought I have, and to-dos I jot down by providing a few considerations and links.
Note: Each prompt used ChatGPT 4o.
Design
If I were to design an app that helps people remember ideas, save links, and track to-dos, the design would lean towards simplicity and clarity.
领英推荐
The home screen can be a list tags generated by the LLM, and a tap on a tag drills in to reveal all items associated with the selected tag. Users can customize a tag’s icon and color, but one can hope that the LLM can also determine which SF Symbol is ideal for each tag. The plus button is the primary action and it launches create mode.
Similar to the ChatGPT app, the focus when creating an item is the blinking cursor. Hidden behind a [+] button can be additional options like uploading a file, uploading a photo, recording audio, or taking a new photo with the camera.
Sending a message to the LLM leverages the same chat paradigm that every AI company is currently using. I truly believe that chat is not the future of interacting with AI. We are in the command-line interface era of AI, and someone will invent the GUI of AI soon.
After the user taps the submit button, the prompt is executed and a response arrives. Of course there is an intermediary step since the LLM requires a few seconds to ponder, but I’m skipping that step here. The response, as demonstrated in the experiment above, includes the category (“Idea” for example), tags, things to consider, and helpful links. The user can choose to save the item here with a tap on the “Add” button, or they can augment the item by continuing to message the LLM.
After an item is created the user can archive or delete it along with a few more advanced options hidden behind a “More” button. The goal is to allow the user to edit the item by removing or adding additional information. Perhaps editing simply reveals the keyboard, or edit mode can be similar to create mode by launching the chat interface again.
Start with the problem
AI is infiltrating every piece of software at the OS level, apps, and conversations. When considering what to build it’s so easy to start with “AI!” and work backwards towards a problem. As Steve Jobs famously said:
I’ve always found is that you've got to start with the customer experience and work backwards to the technology. You can't start with the technology and try to figure out where you're going to try to sell it.
Companies and developers need to consider how to bring tangible benefits to customers using AI. Don’t just imagine cool things AI can do. Help me save time. Help me be more efficient. Help me get away from the computer so I can hang out with my kids. Writing notes, tracking to-dos, and reading documents are activities we all do on our phones and computers. Let’s start experimenting with enhancing them with AI. They might be boring topics, but people will love it.
Senior Product Designer at Salesforce
8 个月Interesting! Love your take on prebaked prompts based on input type. I wonder if Tags should be there on the menu ask for easily going to back to a specific or related convos?