AI. Where to start?

Cutting the Gordian knot

My solution to this was to force myself to do something by signing up for some online training. Even though I am not a project manager I signed up to HustleBadger and the Build with AI course run by Ed Biden, it was four weeks long, with two sessions each week and helped me get my head around AI a bit better.

Learning by doing

During the course of my studies with Hustle Badger I made a whole set of little prototypes to test out and understand the limitations of what AI could do. It was very much learning by doing, and each week Ed focused us on a different aspect of AI.

Week 1

Explored AI agents. I think if you are a project manager, a bit more of your time may be spent using these tools as they seems to supercharge automation.

I explored relay and used it to imagine I was Nick Fury and building a different team of super heroes for different problems.

Not a real world scenario but I wanted to explore how AI understood a data set within an email account and could then filter out the relevant data into a Slack feed. So I used ChatGPT composed a top ten (like Top Trump playing cards) set of superhero profiles and emailed them to a dummy gmail account. Then responding to an instruction in Slack, Relay would serve up a team for Colonel Fury in Slack .

Week 2

Started to look at prototyping and picking the right tools. Looked at Replit, Lovable, Bolt, Figma Make, and Cursor. I had the most joy with Lovable.

I took a prototype for a FX Market orders widget and backwards engineered a 2229 word prompt for the app based a loose prompt framework that I found on the web.

My first revelation was the front end worked well enough to make a great prototype for testing, the currency conversion functioned, the user could input amounts and save orders. it was much better than any prototyping flow you could make in figma. Which would take at least a day to design. Admittedly the prompt took a couple of hours as I wrote it myself

The second revelation was the quality of the copy writing was amazing, really clearly explained each stage.

Finally the bombshell was Lovable wrote some copy, unprompted, explained that a by setting up a rate limit and then a budget rate within a timeframe, the user is actually setting up a OCO order, ‘once cancels the other’. I had written the prompt so that this was not included lovable ‘worked it out’, ‘deduced it’ on its own.

I also got ChatGPT to optimise the prompt for lovable and ran it through lovable again to see if it made a difference. It was a bit better. I finally asked lovable to make an FX orders tool with no further information.

My conclusion was on my first attempt at a prototype based on something I have already designed was, Lovable made a better prototype than I had in figma because the functionality was better than figma’s and it was a better copy writer, but aesthetically it was a bit off.

However if I left it to its own devises with a short prompt Lovable went off with pastel coloured crayons and the results was over complicated and over wrought.

So as I designer I didn’t feel completely redundant.

If you wanted to test a concept and weren’t worried about pixel perfect look and feel Lovable was great. I did also mess around with Replit but got through credits very quickly. Lovable seemed a bit more forgiving.

Week 3

I think we the previous week I had sort of jumped ahead as for Week 3, we were studying workflows for building with AI prototyping tools, understanding prompts for building, and the what we discovered could be the never ending loops of debugging.

Years ago I had brought a Hobinichi medical diary ‘Dear DoctorS‘ and it was all written in Japanese, so I couldn’t understand it, so I never really used it. I wondered if I could make a digital version, mobile first, and style it with a sort of mid century minimalist modern aesthetic. It seemed like it could be a more informal version of the NHS App.

I was also thinking about how people are willing to trust an A.I. machine with personal information. I wonder if because conversational artificial intelligences are designed to tap into peoples tendency to anthropomorphise. User’s often personify the A.I. making them happier to share personal information with. Also unlike a real person the user doesn’t have to worry about the AI’s reaction, they aren’t real, they don’t judge.

Firstly I wanted to understand the diary, and I used Claude to translate scans of the pages. This gave me an understanding of the diary structure, purpose and intention.

Then I compiled the translations into one simple text document and got Claude to turn into a high level spec document

To baseline Lovables ‘ability’ to create a digital health diary I created a simple prompt to see what it would do.

‘Interactive Health Diary Mobile App – Development Prompts
Project

Project Overview:
Create a mobile-first web application based on the Hobonichi Health Notebook, allowing users to track their health data over time.’

It created a very complicated interactive dashboard in what I have come to consider ‘notion pastel’ which seems to be an AI default.

Next I used the full prompt created by Claude. Lovable generated a first pass at the app. As it was a complicated diary with different sections Lovable didn’t get the information architecture correct and it was all a bit of a mess.
It took alot of additional prompting to get lovable to create anything close to the original diary’s content structure.

I wanted to experiment with look and feel and was thinking about a very simple mid century medical look and feel in part inspired by Geigy packaging.

Despite uploading a set of visual designs to Loveble and also exporting a number of layout examples from figma, Lovable didn’t interpret them correctly.

I chatted with lovable and it recommended exporting from figma using tailwind. This didn’t help that much and I ended up burning through tokens.

I concluded that although it was great at functionality it was very very hard to prompt it to a convincing look and feel. Kinda felt like the prompting was getting to Linus levels of complexity and all I wanted was Charlie Brown simplicity.

There is definetely a tipping point where prompting becomes counter productive. A conversational interface is also very long winded to read. I didn’t feel Lovable was ‘design ready’.

Week 4

Fairly complicated stuff. If you were wanting to build something for real and wanting to charge, the course covered implementing a stripe integration, backing up versions to Github, and then we looked at an IDE like Cursor. Whilst I completed the tasks, This was probably at the limit of my technical know how.

In experimenting with a stripe integration with replit, the AI had the comedy halucination that it couldn’t implement the integration because it was a Project manager and not an AI capable of writing code.

I certainly wasn’t comfortable in an IDE pushing code around because I’m simply not a coder.

Conclusion

The course was a great starting point and got me over my aversion to engaging with AI. It also meant when I discuss AI with friends I don’t sound like a complete idiot.
Raised more questions than answers, but learnt alot.In experimenting with a stripe integration with replit, the AI had the comedy halucination that it couldn’t implement the integration because it was a Project manager and not an AI capable of writing code.I certainly wasn’t comfortable in an IDE pushing code around because I’m simply not a coder.