How I Use AI as a Solopreneur

The jobs for which I have been using AI as a solopreneur.

I've come across several posts or comments that mention how AI helps them with coding, learning or automating some work but it’s hard to tell reality from hype with technology these days.

I'm curious about learning how different people are using AI to help them

• save time

• be able to do things that weren’t possible before

• build applications

• be better at teaching

• knowledge management

• learning skills or hobbies

• pretty much anything useful

So I thought of writing a post about how I use AI while running Unwrangle solo. If I get positive feedback I will write posts biweekly-or-so on the jobs for which people are using AI.

1. Writing code

I'm gonna dive right into it. At first, when chat GPT came out I was pasting my code into it and manually adding its responses to my code. Pretty soon I came an across a project on HN that allowed using the API with any repo and I did use it for a while before I switched to using Cursor.

Cursor felt like a god sent because they have an apply feature that saved me the effort of having to copy and paste code modifications made by GPT or other model. The in-line diff style rendering of the modifications makes it a lot easier to use AI for coding, at least for intermediate programmers like myself.

A simple example of the inline diff style UX. I can choose whether to accept or reject every single modification that GPT proposes for a prompt.

Cursor has been quick to allow users to access the latest models from all major providers and now lets users choose between a wide range of models from Claude, OpenAI and Google. At first I would always stick to using GPT-4 for most requests. This changed after Claude released Opus 3. Since then I would save the 10 free Opus 3 requests per day that Cursor offered, for longer and more complicated tasks. I noticed that unlike GPT, Opus would not ask me write the # rest of the code myself and it would also make fewer mistakes. That made it preferable for tasks that required more nuanced logic application and needed the AI to not be lazy.

Nowadays, all the models are improving at a fast pace and I use different models for different kind of problems.

• GPT-4o is nifty with pattern based stuff—think user input based exception handling, regex, pretty much any boilerplate.

• Claude Opus 3 is still the big daddy for me when it comes to not being lazy and handling nuanced or non-obvious cases.

• Claude Sonnet 3 seems to be the best at building UIs. Their artefacts application is A LOT better than Cursor at building UIs with images for reference.

• The long context chat option with Sonnet 200k or Google’s Gemini is useful for tasks that involve working with static files.

So that's my experience with using AI for writing code. This is the main job that I've been employing AI for in my life and I couldn’t be happier about it. If you also code with AI, I would love to learn about your preferred way of doing this and what your experience with different models has been like and how you have been choosing which model to use for what task.

Artefacts vs. Cursor: Here’s a comparison of a Table of Contents UI component that I could create using Claude Sonnet 3 on Artefacts vs. GPT-4 on Cursor with an image for reference. Although Sonnet 3 is available on Cursor, it doesn’t allow attaching an image for models other than GPT-4 at the time of writing this.

Table of Contents that Claude Sonnet created for me via Artefacts using an image that I provided in the prompt as a reference.

Table of Contents component that GPT-4 built via Cursor with the exact same reference image and prompt.

So as you can see Claude Artefacts is far superior to Cursor with GPT-4 at building UIs.

2. Search & learning

The 2nd use-case for which I have been turning towards AI often is search. Like many users, I had grown frustrated with the reducing quality of search results by Google long before GPT was a thing. I had tried using DuckDuckGo in the past but I switched back when I realised that I was doing myself disservice. I had resorted to searching on HN, Reddit, LinkedIn (for companies), ListenNotes (for podcasts), etc. before trying my luck with Google.

Nowadays, these sites have also been useful for me:

1. Devv.ai - I found out about this on HN and I've been using its Github mode to ask questions about open source projects. It has helped me save time while trying to get up to speed with an open source library. Something that I gladly welcome.

2. Aider - I've been using this open source tool to take a deeper dive into open source code before trying to change it for my purposes. I ask it to tell me what the code in each file in a directory does and it helps me map the structure of a repository better and more quickly. I tried using it with my own projects but it has a tendency to make modifications to the code even when I don't explicitly ask it to. So I need to remind it repeatedly to not make modifications to avoid unnecessary changes to my code.

3. Exa - I've been using this neural search engine to find libraries on github and for finding alternatives for tools I'm using.

4. GPT-4o on the phone - I sometimes use the voice mode to ask GPT questions in a conversational style. I actually do enjoy doing this with GPT but it always makes me feel a bit creeped out too.

5. Self-hosted instance of Perplexica - I have been using Perplexica as my primary search engine for web search related to technical topics and it's been surprisingly useful. For e.g. when I used it to ask “if Cursor uses GPT-4 requests quota for Sonnet”, I got a better answer than Google or Perplexity—which is a big deal as it’s not only better but also open source! It uses SearXNG to source the search results and GPT-3.5 Turbo as its chat model provider. SearXNG is a free internet metasearch engine which aggregates results from more than 70 search services. Perplexica can also be used with other chat model providers via Ollama & Groq. More info can be found on the repo linked above.

Self hosted instance of Perplexica which I’ve been using as my primary search engine for at least a few weeks. It uses SearXNG for getting the search results and GPT-3.5 Turbo as the chat model.

It feels amazing to use an open source search engine and for it to be this performant. This project deserves an entire post on it. Let me know in the comments if you’d like me to make my instance available or do a post on how to set this up.

3. Documentation

This is something new that I recently started doing with AI. I need to manage a lot of services on different servers for my product Unwrangle. I maintain a wiki on Notion where I put set up and maintenance instructions for all the different technologies I rely on. AI has saved me time here too!!

I now save the contents of .bash_history (a file that saves a record of the commands you make on a linux cli interface) after performing any operations I need to document to a txt file. I then upload this file to Google’s AI Studio and ask it to write me a confluence style instructions document that I can refer back to when I need to repeat that operation in future. Writing docs is important when working with technology but it’s also time consuming and feels like grunt work. So this is my new favourite job that I like to automate with AI.

4. Writing

For a non-native speaker my writing is not too bad (I wrote this without AI). But writing emails for Sales and Support is a different animal and can be frustrating at times. It's challenging for me to find the correct voice between sounding indifferent and pushy or callous and needlessly flowery.

Claude has been useful here. In particular their new projects interface. I pass it relevant custom instructions and I noticed that it makes a huge difference on what it generates. Until recently I never used AI for writing emails because I found them to be too generic, pushy or just creepy. But this has changed with Claude's projects. I give it some context, the email I’ve written myself, sometimes the conversation trail and the changes it suggests have actually helped me drive sales.

As an engineer with no formal sales training I tend to be too straightforward. However, selling does require a slight nudge or an ask so-to-speak. And Claude has been helpful in writing emails that help me strike a balance between too plain and too pushy.

So these are all the ways that I have been using AI to do more. I have been beyond excited about the possibilities this simple and powerful technology gives us all while being weary of the hype. I intended to keep this short. If you read until all the way here I hope it was worth your time.

If like me you’re curious to learn about the jobs that people in different fields are automating with AI, please subscribe!! This will motivate me to put more effort into researching how people have been using AI to help them do things that weren't possible before.

This is my first long form blog post ever. I'd really love to hear about how you have been using AI or the jobs for which you have been using it. You can reach me at raunaq@unwrangle.com.