How We’re Using AI to Save Time and Write Better Code

The AI revolution is producing a ton of great tools to assist developers and I wanted to write a post about how we’re currently using those tools here at Harmonic Northwest. I think it’s probably easiest to explain through some real-world examples which I’ll go into below.

AI-Assisted Development Example #1: Scaffolding

A client needed a complex branching form setup where the question fields that would show depended on answers to previous questions. Additionally, they wanted the data for the form to be in an easily editable data file.

Normally I might do something like this using WordPress’ Gravity Forms plugin via the plugin’s conditional fields in the admin area, but in this case that wasn’t going to work because we needed a single data object that would power the logic.

So instead I fired up Claude Code and started writing a requirements document. I gave Claude a paragraph or so describing what I wanted (JSON object for data, conditionally show/hide questions, store answers in hidden fields, validation, etc.) and had it write a formal requirements doc. I looked this over, gave feedback, and we iterated on the doc for a while.

Once that looked good, I asked Claude to review how the current codebase for the project worked so that it understood the current state of affairs. Then I let Claude loose on the code. It produced some new JavaScript, CSS/SCSS and PHP code with all the structure needed to power the functionality. I updated the data file with real data (this part took some human parsing to get right) and the functionality mostly worked great. We did some more iterations after testing with real data, but it wasn’t long before the project was ready to ship.

Scaffolding complex projects by hand can be super time-consuming, but for the AI it was much more efficient. Claude ended up producing several hundred lines of code, which, if I had done completely by hand, probably would have taken the better part of a week. Instead we had a first draft up with a completely working form and real data in under a day.

AI-Assisted Development Example #2: Working with Complex Data Structures

The JSON data object mentioned in Example #1 ended up being somewhat complex. We ended up with a couple dozen questions and a lot of metadata around the conditions under which those questions would show. There were also a bunch of display options for each question such as label, help text, and field format (simple text input, large text input, checkboxes, etc.). This large amount of data made the data file a little tricky to edit if you wanted to, say, add a new branch to the logic that showed a series of new questions.

But navigating these kinds of updates with the help of an LLM was very simple.

Instead of manually editing lines of code in the JSON file, I could ask Claude to do what I wanted, and for Claude this was a very simple task. For example, I could say something like, “Can you remove the third answer for the question titled [title] and replace it with this question [question] and have it be a multiple-choice question with responses [A], [B], and [C]?” Claude could handle this request in seconds without breaking a sweat.

This dataset wasn’t all that big—just a couple dozen question entries. But the same principle applies to larger datasets, like a spreadsheet file in CSV format with several thousand entries.

Another example along these lines that I did recently: I exported a bunch of WordPress data in standard XML format and then had Claude parse through that data to produce a spreadsheet-compatible CSV file with just the SEO data. This would have taken a while to do by hand, but with Claude’s help it only took a couple minutes.

AI-Assisted Development Example #3: Debugging

My AI assistants have become essential for debugging code. I generally use them for debugging in two ways:
analyzing the codebase and
producing useful log messages.

I was recently having a problem with the shipping calculation portion of a checkout process for a custom Shopify theme. I didn’t have a lot of useful information coming back in the browser or server logs, so initially I just told Claude what was happening, pointed out a couple files I thought were likely involved, and asked it to assess the code to see what might be happening.

Sometimes in this scenario Claude will propose a solution right away and the problem is solved immediately. In this case, we needed more data. I asked Claude to write some debug logs that outlined several data points that were potentially contributing to the issue. After doing more testing with the new logging enabled, I was able to identify the root cause of the issue. I pointed this out to Claude and suggested a fix and Claude implemented the fix magnificently.

AI-Assisted Development Example #4: Understanding Documentation

I recently was working with an API I’m not terribly familiar with and needed to know how to pass an extra data type that I hadn’t worked with before. I asked Claude to look at the documentation to see if it could figure out how to do it. Claude scoured several sources of information (not just the documentation page) and came up with a pretty good answer. It might have taken me an hour to sort through this; Claude gave me testable API calls in a few minutes.

AI-Assisted Development Example #5: Generalized Knowledge

I’ve been using Claude and ChatGPT to fill in gaps in my knowledge around best practices for SEO, site performance and accessibility. I can oftentimes ask it a question like, “Do I need to add an ARIA label for this DOM element given that its functionality is X, Y and Z?” and it usually gives me an answer that if not 100% correct, at least gets me pointed in the right direction.

A lot of times these questions surface new bits of knowledge that I wasn’t aware of and lead to further conversations that further extend my knowledge.

Adapting to New Standards

This is the gist of how the Harmonic Northwest team is using AI tools today, but of course the tools are changing and improving and new ones are popping up all the time. The way we interact with LLMs and other AI tools is going to change.

We don’t know what the future holds, but we’ll continue to learn and adapt so we can do our work efficiently, integrate the industry’s best practices and produce more robust and plentiful code.