How to think with AI?

Foundations to know

• AI is best at small, specific tasks. Examples include making a page from a design system, making a table with sort, filter, and search, creating a simple design system in code, and writing an implementation document based on screens.
• AI is not good at large, complex tasks. Examples include making a multi-page ecommerce site, making a B2B application with workflows, and writing accurate documentation for implementation.
• We think about AI in terms of a junior designer or developer who does a lot of work, but who is unpredictable and unreliable. You never know when they will produce impressive, mind-blowing work, or when they will produce laughable (yet entertaining) work.
• So we read and view everything that AI produces with a critical eye, and we take only what's useful (many times edit). AI has been producing more useful output than it did a year or two years ago; however, we do not know when it can be fully trusted to work on its own.
• What about productivity? We have written a separate article on this topic. The summary of it within this summary is: AI is a productivity multiplier for an individual, but it barely moves the needle when the team grows larger than 10 people. Productivity gains become even more negligible when there are shared responsibilities and dependencies between team members — for example, when you have 3+ designers working on a product, or a large codebase where 4+ developers need to collaborate.
• The common thread/reason for all of the above are 2 things:
  1. AI output is probabilistic with little or poor judgement.
  2. AI cannot hold and process enough information, otherwise called context to do complex, long duration, multi-session tasks.

How to get the best out of AI?

Brownfield projects

Break down the task at hand into small enough, specific enough tasks. We will put out examples for this soon. If you are in a brownfield project, work with your developers to compact your repo into a bare-minimum code repository that only has your design system and can render a screen when you run it. You play the role of judge of AI output — review everything it produces, edit where necessary, and hand it over to the person or people responsible for taking it further. This assumes you know how to read and review code for design. It also assumes your developers are able to fill any gaps you may have missed. There is a tremendous amount of learning that needs to happen on both the design and development side for this to work.

Greenfield projects

If you are in a greenfield project, you have the opportunity to build your design system in code from the start. Keep your design system lean — fewer than 40 components. This is the best way to get good output from AI and to be able to use it for more complex tasks. You can also use AI to help you build your design system in code, but you need to be careful and review everything it produces with a critical eye. We will put out examples for this soon. Your role as the judge of AI output is still relevant here, but the amount of context you — and AI — need to process will be much less. Once again, there is a tremendous amount of learning that needs to happen on both the design and development side for this to work.

Back to the productivity question

Essentially, when you reduce the amount of context that AI needs to process, work that would take two or three people can be done by one person with AI. If you are able to compartmentalise the work such that dependencies and shared responsibilities are minimal, and the trust between the people working on the project is high, you get great productivity gains with AI. At least, this is our hypothesis.

Our skepticism of the skeptics

The skeptics will likely say things like "this isn't going to work," "this is going to be a nightmare to maintain," or "this is a laughable suggestion." In my time as a designer — and more recently as a designer-turned-developer — I have seen small teams accomplish great things when they understand the fundamentals of the product they are building and the craft of product-making itself. Examples include Zerodha, WhatsApp, and Mailchimp, all of which achieved great success despite being very small teams. As AI becomes more mainstream, more teams that understand how to use it effectively will be able to compete with larger teams that don't.