How we use AI to build products faster than ever

Whew, it has been hot over here in Europe! I've spent this week out at my winery in Priorat, Spain, where the highs are around 100ºF. We've had a great growing season so far and this vintage is looking really good. Plus, one of our previous vintage wines was just awarded a Silver medal and 92 points from Decanter, which I'm very proud of!

Last week I talked about what I'm doing with Beta Acid's team and why. This week I'll dig a little more into the how.

It's a bit of an oversimplification, but there are three use cases for how we are using AI to help us build software:

1. Product discovery, requirement gathering, and documentation creation
2. Assisting developers with code generation and technical architecture
3. User-facing AI-enhanced features

Before we can talk about how we enhance products with user-facing AI features (which is a fascinating topic on its own), we have to actually build the software, so we'll focus on #1 and #2.

The most impactful AI workflow we've adopted is Spec Driven Development. SDD is an iceberg; it's fairly easy to understand but mastering everything below the surface takes a lot of time and understanding. It's compounding. SDD takes vibes and replaces them with a rigid, reliable structure. Like vibe-coding went to university.

You may have already heard of an AGENTS.md file or a CLAUDE.md file, which many AI code generation platforms like Claude Code, Cursor, and Codex use. Essentially these files are a set of instructions that the AI agent is required to follow when they are performing code generation tasks. This is useful, but simplistic and primitive.

We take this a few steps further and build an entire library of instructions for our AI agents.

These Spec documents generally fall into a few categories:

• Incredibly thorough engineering best practices and architectural guidelines that we've refined from building dozens of products. We often reuse these for every new product, improving them every time.
• Product requirements documents. Typically one or two that describe the product as a whole and then one for each of the major features or components of the product.
• AI Agent planning documents that define technical implementation requirements that are created as the product is built.

In a typical (ie. complex) product we build these days, we might have more than 25,000 lines of instructions across our Spec files. These are way too complicated to share here, but let me know if you want to see examples.

Many people are under the illusion that you can build complicated products in "one shot", but we have found the best way to build using SDD is more analogous to how we traditionally build software: an iterative approach that keeps building on top of and refining itself.

We start by defining exactly what the product is: its many features, capabilities, and technologies. This process often takes us several days to complete. From that Product Spec we can build a technical architecture plan and use that to lay the foundation. From there we build the product from the ground up, piece by piece.

AI is the great enhancer. It doesn't magically give you skills you don't already have. If you weren't a doctor before using AI, you're still gonna be a shitty doctor with AI, too.

What AI does do is enhance the skills you already have. Because our team comes to the table with decades of professional experience in these areas, AI tools allow us to be even better Product Strategists, Designers, Engineers, and Testers.

This is how we create, refine, implement, and review excellent Spec documents that actually work.

It's how we built our latest product, Nudge, the task-oriented CRM that I use every day.

That's it for this week. I've got some wines from our 2025 vintage to put labels on!

Cheers,

Ryan