I am building Answer HQ with AI - here is what I've learned so far
Conclusion: Software development is never going back
I haven’t coded professionally since 2018 (that’s 6 years, a lot of time in the software development world), yet I was able to build Answer HQ over the past few months.
P.S. We are currently personally onboarding folks to Answer HQ. If you are a small business or startup, and want to reduce your customer support questions by 80%, reply to this email or join the waitlist. It’s still (obviously) early, so we are more than happy to listen to feedback!
👉🏻 Here’s how I did it and what I learned so far
Cursor and Claude Sonnet 3.5 are workhorses
These two tools have turned me from a washed and rusty dev to a slightly less washed and rusty dev.
What I noticed about using AI for coding is that the emphasis of knowing syntax and language idiosyncrasies became less important, but thinking about system design, architecture, extensibility, and code flexibility became much more prominent. When I am able to purely think about the high level, I am able to make better decisions
LLMs still struggle with UI and frontend
Claude Sonnet 3.5 is the best all-arounder for coding both backend and frontend, but all LLMs currently still struggle with building /good/ user interfaces. Reason? It can't see the physical UI it builds - even sending a screenshot in isn't that effective.
You need to know your React/Tailwind/etc extremely well to properly steer it. I am not a frontend person, so I took a lot more time on frontend in steering it to give me an ideal output
AI-powered UI builders like v0.dev are okay starting points
Sure, v0.dev lets you use prompting to create a pretty good UI to start with, and you can even easily import the code through NextJS commands, but it kind of falls apart when you want anything more customized than the basic Tailwind/NextJS "look" (if you've worked with NextJS/Tailwind, you know what this "look" looks like)
It has potential, for sure, but UI/UX still has a long way to go for LLMs IMO.
OpenAI’s o1 series is pretty good, but it’s not a daily driver yet
The reason I use Claude Sonnet 3.5 and call it a dailydriver/workhorse is because while it's not *the best* at coding or large context, it does everything generally really well. It's also way faster than o1 + fewer rate limits.
The speed of o1 really inhibits development time.
BUT, it's 32/64k respective (preview vs. mini) output context limits is an absolute game changer. It's so good for large code bases + refactors. I’m very excited to find out how Claude Opus 3.5 with higher output context limits will fare.
You still need to know how to code
LLMs currently are great at MVPs and demos, but to go anything beyond that, having a background in coding is *still* needed
There are a bunch of bugs and features I developed where I have to make active decisions (especially on architecture) on what to do next, and an LLM wouldn't have been able to give me that answer. For example, how I'm storing my user sessions, settings, etc.
Software development will never be the same
This is going to anger some nerds, but coding won't go back to the way it is. Even if you don't think AI is transformative, at least give AI-powered coding a try. The least that can happen is you get to see what it's good and not good at - this information is crucial.
Building, for now, is still hard. Humans, for now, are still required.
Building a product/company/business is still a very human process.
The majority of building something is beyond what AI can (currently) replace - things like being persistent, working long hours, finding emotional support when you're down, making connections with customers and people, cold emailing prospects, writing good copy (I use LLMs daily and they *still* suck at writing), etc.
This doesn't mean putting off learning about LLMs - still use it. Learn what it’s great at. Learn what it absolutely sucks at.