The highs and lows of building a micro-SaaS with my pal Claude

I’m writing this because a lot of people feel uneasy about AI. As someone who slurps heartily from the LLM firehose (for now at least), I completely understand these feelings.

I oscillate between “this is amazing” and “this is very disorienting” on a regular basis. Despite my mixed feelings, AI now augments most of my design, coding, and ideation work. Given that, I feel qualified to share some opinions formed while building my own software, solo.

Where it began

4 months ago I left a good, stable job in legal technology. By most accounts this was the behaviour of a crazy person, especially given the current economic climate.

I don’t deny that. It was a yolo but I felt compelled to do it because of the incredible change occurring in the world of AI. I was noodling with ChatGPT and Claude in the year leading up to this point, but the capabilities were accelerating and I felt inspired to act.

I had identified a specific set of problems related to condominium maintenance that really bothered me. These problems consumed my time as a volunteer and my money as an owner. Naturally, as a designer, I thought “software can fix this”.

At any other time in history, I would not have tried to build this product. Not because I couldn’t write code, but because it would have been impossible to do so within a reasonable time for one person. But now, with Cursor or Claude Code, the calculus has changed. I thought “now or never”, so I committed and started building.

4 months later

I’ve built a production ready, multi-tenant, responsive web application called CondoWrangler. It’s not mission-critical software running spaceships but it’s not a hobby project either. Email to ticket ingestion (think Zendesk), asynchronous task processing, AI integrations, image annotations and optimization, etc. A ton of work went into building this system so I could commercialize it and trust it to run reliably.

And this doesn’t even touch on the marketing and sales work done in this time. It’s truly remarkable how much ground one person can cover in such a short period with these tools.

“Wow, Geoff. we get it; AI is neat” I hear some of you say. Well, maybe…

Where AI has genuinely helped

Design

I design exclusively in the browser now. I only occasionally use Figma to create or modify images. This is the greatest shift in how I practice design to date. I imagine the same will be true if working on native applications. The final medium is where I want to work.
Image generation. This can be hit or miss but you can get pretty far with the right prompting. That said, AI generated images still have a smell and they need to be used judiciously.

Development

Helping brainstorm app architecture. This was a big unlock because it exposed me to different patterns that I wouldn’t have known about without considerable research.
I spend very little time writing boilerplate code anymore, especially on the frontend. I used to enjoy this work but now I happily point Claude at a css file and let it rip.
Writing non-critical features with little to no involvement outside of prompt development. There are standard CRUD features where I touch zero code. I prompt, review, send a PR to Google Code Assist, which I then feed back into Claude Code. Assuming there is reliable precedent in the codebase and I can perform reviews, just let the robot do it.
Documentation. I will likely never write code documentation again in my life. The AI is so good at it.

Marketing and research

Using agents to perform market research. This has been hit or miss but in general it’s a win. The alternative is I would have to do it all myself. This at least provides me with a foundation to build on.
Having the AI review my marketing site and optimize it for search.
Workshopping digital marketing techniques. I’m not a digital marketing expert, so having an AI give me the lay of the land, and potential paths to explore, has been helpful.

Writing

I use AI to review and edit my writing. As an editor I think it’s great but I will never let these things write on my behalf (thoughts on this below).
It can be very useful as a sounding board in helping structure things like pitch decks or scripts. With the right guidance, it’s good at trimming the fat from my messaging.

Where AI is not helpful

Surprise! Most of the hard, important work still exists.

Design

I still have to identify the right problems to solve, the right workflows, the right copy, etc. I still have to talk to customers and understand their context. I am more convinced than ever that AI cannot do this work competently. Perhaps for commodity apps it works, but anything with specificity or for non-technical audiences good friggin luck. I don’t buy it.

Development

Agentic coding encourages you to close your eyes and let go of the steering wheel. This may feel exciting in the moment, but I can’t shake the feeling that it’s going to create chaos down the road. I make a concerted effort to understand my codebase even when agents are writing features for me. When something does inevitably go wrong, I won’t be at the mercy of the system.
Because of how frictionless agentic coding is, it’s easy to take a “shoot first, ask questions later” approach to building. On more than one occasion I’ve wasted a fair amount of time (and tokens) on features that should never have been built. Previously the cost to writing code was high, so I was forced to be thoughtful. Now any bad idea that crosses my line of sight can be manifested in minutes.

Writing

I will never let LLMs write for me. Edit or brainstorm, sure. Write, never. Three reasons:

AI prose has a smell to it that immediately turns me off.
AI’s ideas are an approximation at best; I’ve never had it produce writing that captures what’s in my head.
Letting AI write for you is a slippery slope. As Paul Graham remarked: Writing is thinking. I’m not keen on delegating my thinking to a GPU in California.

Business development

Perhaps the most important domain of all. The source of income and survival for the company. I suppose I could set up a Clawdbot and have it cold call people on my behalf, but I’m not a monster! And what percentage of prospects are going to respond positively to that?
At the end of the day, CondoWrangler lives or dies based on my ability to connect with people and solve their problem. I can’t just point an LLM at this problem and expect anything worthwhile to come out of it.

Clarity of mind

AI allows me to work at an accelerated speed of which my brain is not accustomed. This sometimes creates a mental fog that I’d previously only experienced after long periods of intense work. It doesn’t happen all the time, but it’s a recurring pattern I’ve noticed that I’m now on guard against. I don’t love it.

The one thing I know for certain

This has been the most exciting but unsettling period of my professional life.

So much possibility and uncertainty crammed into a single space in time. Very weird.