Workflows for the Win (for Developers)
The case for workflow-driven AI software development.
This post assumes some software development experience. If you’re new to AI software development, you may prefer this version: Workflows for the Win (for Non-Developers).
At CodeYam, we’ve become strong believers in AI workflows. They are a core part of the CodeYam experience. Workflows provide a robust and consistent development experience that gives us confidence that the code being written is not introducing technical debt that will make it harder and harder to make changes to the codebase in the future.
So what are workflows exactly? How do they differ from other AI tools or techniques such as skills?
We define a workflow as follows:
A workflow is a step-by-step process with automated audits and check-ins at the end of various steps.
This differs from AI skills in that a workflow breaks down a complex challenge into many steps and delivers the instructions for each step only after the previous step is completed.
You could do this with skills by having many skills that you apply as you move through the challenge, but a workflow seeks to automate this process and might have additional checks and audits at the end of each step. This keeps each step in a workflow small and easy to understand.
This is not complex technologically, but deconstructing a complex task into smaller sub-tasks is an often overlooked strategy when dealing with complex challenges. The book The Checklist Manifesto provides ample evidence for the value of taking complex processes, breaking them down, and creating checkpoints for each well-understood stage of the process. Steadily working through each checkpoint ensures that nothing is forgotten and frees up mental capacity for the more challenging, creative parts of the work.
Making a change to a codebase, such as building a feature, is a complex process. You don’t want to just write the code sufficiently to make the change work. You want to write tests, ensure the code is well-organized and easy to find, ensure that nothing else breaks unexpectedly from the change, etc. This is in addition to making sure that the change performs as expected and that you like how it behaves once the change is in place.
A primary goal of software development, beyond building the product, is to ensure that you can make changes to the product quickly and effectively even as the complexity of the product grows. This is no easy feat.
Software explodes in complexity as you build out additional features. Ensuring the code remains well-tested and well-organized as you quickly iterate on the product has always been a challenge for software teams.
AI introduces new benefits and challenges toward achieving this goal. AI can go very fast, but, without additional context and effort, is not likely to keep your code in a highly maintainable state as it rapidly builds the application.
Workflows can take the large language models that exist today and guarantee a more reliable process for keeping a codebase maintainable. Workflows can help LLMs do the right thing consistently regardless of the AIs ability to think through such a challenge itself. And since AIs can work so fast the extra effort to ensure the codebase remains maintainable is not a significant cost.
A simple workflow might be something like red → green → refactor, which has been a mantra of test-driven development for decades. Even this simple workflow was hard for human engineers to do reliably. The temptation to just code until the code works and call it a day without thinking about how maintainable the code is has been a constant struggle for engineers as long as code has existed. AI changes all of this.
AI has no problem following a strict workflow, regardless of its length or how tedious each step is in the workflow. The AI doesn’t get tired or bored the way a human developer might.
At CodeYam we have a workflow that, at a high level, looks like this:
Plan
Prototype
Demo
Deconstruct
Glossary
Present
Save
The goal of this workflow is simple. Start by planning the feature. Then rapidly prototype and demo the prototype to ensure that there is a shared understanding of the desired outcome. At this point, it’s easy to iterate on the code to make changes if needed.
Once the prototype is approved, we can deconstruct the code to ensure all functionality is organized and test-driven into small, well-defined functions with clear boundaries that are well-tested. This essentially creates a well-constructed API within the codebase.
That API can then be documented in a glossary that records every function, its usage, and how it is tested. This glossary serves as the organizational step so that every aspect of the internal API is easily discoverable, encouraging reuse, and as an audit itself, ensuring that every entry in the glossary is recorded properly, matching the code, with tests to demonstrate its functionality.
We then break these fundamental steps down even further into a workflow that has over 25 steps currently (and is constantly evolving) with a number of sub-tasks at each step. Many of these steps have audit gates that run static code analysis to ensure that the AI has not forgotten anything and that the system remains tech debt and drift-free.
This workflow allows us to kick off any feature knowing that the AI will check in at predetermined times to allow the user to confirm the work is being done as expected, with robust steps and audit checks along the way that essentially guarantee the work is done in an organized and reliable manner. This creates a nice rhythm to the work as well as you can easily get a sense as to how far along in the workflow the AI is and when it is likely to complete the work.
Is this completely necessary? Will AIs be able to handle this on their own and reliably ensure that the codebase is always clean, well-tested, well-organized, and maintainable? Possibly. Right now, all of the top models we have tested struggle to do this effectively. Even if they were able to without a strict workflow, though, there still may be advantages to workflows.
Much as the research in The Checklist Manifesto revealed, checklists are valuable tools in any complex work, especially work that is complex and creative.
Why ask any intelligence to keep in memory dozens of to dos when it can externalize that work to a tool? And why not provide that intelligence with tools that help it double check its work or that can automatically handle tedious aspects of the workflow?
We’ve written about this before in To Tool Or Not To Tool, but the idea that an intelligence would avoid using tools that make its life easier because it is intelligent enough to do the work without the tools is ridiculous.
If anything, a sign of higher intelligence is that it uses more tooling, not less. The only situation where it makes sense to avoid a tool is when the task at hand is so easy that even engaging with the tool is a waste of time.
So the better question becomes why wouldn’t you use workflows and tooling to make the job an AI is trying to accomplish easier and more reliable?
It’s unlikely that the challenge of maintaining a robust, clean, well-tested, and well-organized codebase will become so easy for AIs that tooling is not helpful. Tests themselves are tooling in this same manner. So you’d have to believe that AIs would be so powerful as to not need tests at all to completely avoid the benefits of tooling, which seems highly unlikely.
Either way, right now we are seeing great benefits from workflows and the accompanying tooling we have created. They work with any technology stack and at any level of complexity. They give us a lot of confidence that no matter how complex the application we are building gets, it will continue to be easy to make changes to it.
This relatively simple technique has delivered the initial promise of vibe coding, where you can express your intent in plain english and expect a well-constructed code base and application to be developed without any further support, far more effectively than working with a raw AI model.
Don’t take our word for it. Try CodeYam yourself. It’s free to use and takes just a few minutes to get started. Simply follow the instructions on the CodeYam homepage.
If you’re building something more ambitious, are new to building software, or want help establishing AI-native workflows for your team, get in touch. We’d be happy to help you get started.
Our mission is to empower individuals and teams to build beautifully designed, highly maintainable software with CodeYam.
 | A guest post by
|