How We Built the First Autonomous AI Coach for Creators

We're a bootstrapped SaaS company who went all-in on AI during the past year. Today we launched Heights AI Coach.

This isn't the first AI feature in our product, but it is the most complex. I'm going to break down our process for building with AI because I feel there is so much potential for indiehackers to build autonomous agents like this for their specific verticals and industries as we did.

Heights AI Coach is what we consider an autonomous agent. If you've looked into building with AI, you probably heard someone mention AutoGPT or BabyAGI before.

Unlike most AI tools where you need to continuously prompt them to get them to do anything for you, an autonomous agent can take actions on its own without requiring human feedback or input at every single step. ChatGPT's code interpreter is an example of an agent as well, where to answer a question it will decide to go through multiple steps, including running code, before it gives you a final answer.

I could go into more detail, but let's talk about how to build.

Here are some key points that we did or do differently:

• Everyone on the team works on prompts (technical role or not). An amazing thing about AI is you don't have to be a developer to build with it. All you need to know is how to write instructions.
• We keep a separate kanban board for prompt progress to distinguish prompts we are working on from other feature stories in our product.
• We test and iterate on prompts to get them working properly once they are built. Most teams don't have a defined process for this yet. It is important to build a feedback loop for yourself, so we have customer feedback (ie: clicking the thumbs up or down button) on AI outputs go directly to a slack channel so our team can understand what to optimize.
• We don't use libraries. No langchain or similar frameworks. What we built has custom requirements that technically aren't that difficult, so it is easier to build without a framework than with it.
• Get users involved early and often! We had our users test all of our AI features before a full general release. This was especially important with Heights AI Coach, being an autonomous coach, we want to make sure that it is continually valuable, and doesn't go off the rails and do something unexpected. We coached some of our users ourselves and compared our guidance with our Heights AI Coach would have recommended, to optimize and refine our agent. We also had users try it before launch to understand how it would work or not in a real-world use case.

I feel too many builders are attempting to make general AI agents, which is very difficult with the tech available today. Specialized agents like what we built with Heights AI Coach are very possible today, and have so much potential.

I am happy to expand on this. Curious who wants to hear more?