Insight

Ask a developer how they use AI in their development process and many will describe the same thing: open a chat window, describe what you want, see what comes back. This works fine for small, isolated features, but not when you want to broaden the scope to functionality that touches multiple software layers.

A mature software development process, integrating AI, is very possible, and at a high level, does not change from the traditional approach. You need to set clear goals, define where the project at a high level is heading, then break down this goal into tasks which agents implement, most importantly test, and repeat.

A toolkit in the form of AI skills greatly helps with this. It allows developers to standardise AI instructions, processes, and assets. Rather than write your own from scratch, the open source community have released various options. The most popular being the skills library maintained by developer Matt Pocock. The AI software development cycle with these skills becomes:

• Discovery + define: Before any code gets written, we use the "grill-with-docs" skill to interrogate the idea properly. The purpose is to get a shared understanding between the agent and the developer. The use of other skills such as "Wayfinder" or "Prototype" can also aid in the definition of the scope.
• Specification: Once we have mapped out the desired functionality, we need to document it as a specification. This is where the "to-scope" skill comes in. It turns the ideas into a proper specification: what's being built and why, blending technical and non-technical detail into one shared target.
• Tickets: Next we need to turn the specification into actionable items. "to-tickets" splits the spec into tickets small enough to finish in a single session, so a large build can be spread safely across many sessions or people. Tickets can be Github issues, local files, or pretty much whatever you want them to be.
• Implement: The "implement" skill works on one ticket at a time. It uses Red/Green TDD methodology to ensure generated code works and does not break previous implementations.
• Code review: Once a ticket is done, the "code-review" skill runs two checks in parallel: does the code meet the team's own documented standards, and does it faithfully deliver what the spec asked for.

So, AI can produce quality code, as long as there is a process to achieve and reinforce it.

Still have questions? Give us a call, book a quick meeting, or send us a message to understand more without any pushy sales tactics.