1 of 27

My AI-Powered

Development Workflow

Swarnika Shakya �Technology Lead

2 of 27

The Reality of WordPress Development

Research & troubleshooting

Debugging

Boilerplate code

Context switching

Reading Documentation

So, where does most of our time actually go during development?

We usually think of development as writing code.

But in reality, writing the code is often only one part of the job.

A lot of our time goes into things like reading documentations, researching how something works, troubleshooting an issue, debugging someone else's code, writing boilerplate, and, of course, switching between different tools and tasks.

And none of these things are necessarily bad. They're part of being a developer.

The problem is that many of these tasks are repetitive and time-consuming.

For example, you might spend 20 minutes searching through documentation just to remember the exact API syntax for something you've already implemented five times before.

Or you might spend an hour debugging an issue where the actual fix takes three lines of code.

This is where I started looking at AI differently.

Not as something that replaces the development process, but as something that can remove some of the friction inside that process.

And that led me to structure my development workflow a little differently.

3 of 27

AI-Powered Development Workflow

Idea

Plan

Generate

Review

Test

Deploy

4 of 27

Idea

Plan

Generate

5 of 27

Review

Test

Deploy

6 of 27

7 of 27

8 of 27

9 of 27

10 of 27

Prompt

Code

Done

Prompt

Code

Run

Observe

Review

✔

✗

11 of 27

12 of 27

13 of 27

14 of 27

15 of 27

16 of 27

17 of 27

18 of 27

What is skills.md?

A Markdown file containing reusable project instructions

Defines coding standards, architecture preferences, and project conventions, security requirements

Reduces the need to repeat the same prompts

Keeps AI responses more consistent across your project

What exactly is skills.md?

It's essentially a Markdown file that contains reusable instructions for AI working on a project.

It can describe the coding standards we want to follow, our architecture preferences, project conventions, security requirements, and other rules that are important to the project.

The biggest benefit is that I don't have to keep repeating those instructions in every conversation.

And there's another benefit that I think is even more important:

consistency.

If I work with AI today and then come back to the project next week, I want the AI to understand the same expectations.

And if multiple developers are using AI on the same project, having a shared set of instructions can help establish a common baseline.

So instead of treating AI as a blank chat window every time, I'm giving it project-specific context.

And that's a shift in mindset.

We're not just learning how to write better prompts.

We're learning how to build better context around our prompts.

19 of 27

WordPress specific skills and references

WordPress plugin structure

Hooks and filters

Settings API

WordPress coding standards

Security practices

20 of 27

agents.md

.agents/skills

21 of 27

AI Beyond Code Generation

Debugging

When people talk about AI and development, the first thing that usually comes to mind is:

“AI can write code.”

But honestly, that's not necessarily where I find the most interesting value.

One of the areas where I use AI heavily is debugging.

Instead of spending a long time staring at an error message and trying to figure out where it originated, I can give AI the error, the relevant code, and the context around the problem.

Then I can ask it to help me reason through possible causes.

And I want to emphasize the phrase “help me reason.”

I'm not asking:

“Here's an error. Give me a fix. I'll blindly apply it.”

I'm asking:

“What could be causing this? What should I investigate first? What evidence would confirm or eliminate each possibility?”

That makes AI much more useful.

Because debugging isn't just about finding a line of code that needs changing.

It's about understanding why the system behaved differently from what we expected.

And AI can be a very useful second pair of eyes during that process.

“And the same idea applies to another part of development that we often don't give enough attention to: documentation.”

22 of 27

AI Beyond Code Generation

Documentation

Documentation is another area where AI can remove a lot of repetitive work.

Developers often know that documentation is important.

But when we're busy building features, documentation is very easy to postpone.

AI can help turn information that already exists in the project into something more structured.

For example, after building a feature, I can use AI to help explain what was changed, document how something works, or organize technical information into a format that's easier for another developer to understand.

And again, I'm not saying:

“Let AI document everything and never look at it.”

The developer still needs to verify that the documentation actually describes the system correctly.

But AI is very good at taking information and helping us structure and communicate it.

So for me, AI becomes less of a “code generator” and more of a general development assistant.

23 of 27

Choosing the Right AI Tool

Planning & Architecture

Claude, ChatGPT

Coding

Cursor, GitHub Copilot

Large Codebases

Claude

Google Ecosystem

Gemini

Local & Private Projects

Ollama & OpenCode

VS Code

Continue.dev

Terminal Workflow

Aider

Autonomous Coding

Cline

There isn't one AI tool that's automatically the best for everything.

Different tools are useful for different workflows.

For example, for planning and architecture, I might use something like Claude or ChatGPT because I want to have a conversation about the problem before I start implementing it.

For coding directly inside the editor, tools like Cursor or GitHub Copilot can be very convenient because they're already sitting inside the development environment.

For a large codebase, context becomes much more important, so I may choose a tool that handles that kind of context particularly well.

If I'm working heavily within the Google ecosystem, Gemini can make sense.

And then there are local and open-source options like Ollama and OpenCode, particularly when privacy or running models locally is important.

Feel free to take a picture of this slide for the reference and Choose the tool based on the problem and the workflow.

“And that also brings us to one of the biggest questions when people start experimenting with these tools: do I need to pay for them?”

24 of 27

Free vs Paid vs Open Source

Free	Paid	Open Source
Great for learning	Better models	Privacy
Limited context	Larger context	Offline
Limited requests	Faster responses	No subscription

The good news is that you don't necessarily need an expensive AI subscription to start.

Free tools are perfectly capable of helping you learn the workflow.

They can be great for experimentation and understanding how to work with AI.

The tradeoff is usually around things like limited requests or limited context.

Paid tools generally give you access to stronger models, larger context, and often faster or more convenient workflows.

And then we have open-source AI.

This becomes interesting when privacy, control, or offline usage matters.

You can run models locally rather than sending everything to a cloud service.

Of course, that comes with its own tradeoffs — hardware requirements, model capabilities, setup, and maintenance.

So I wouldn't frame this as:

Free is bad, paid is good, open source is best.

It's really a question of:

What are you building? What constraints do you have? And what does your workflow require?

You can start small and upgrade only when you actually hit a limitation.

“But regardless of which tool or model you choose, there are some things I would strongly recommend you avoid.”

25 of 27

Best Practices

Avoid

Blind copy-pasting

Skipping testing

Ignoring security

Sharing API keys

Assuming AI is always correct

This is probably the slide I would remember if I were sitting in the audience.

Because as powerful as these tools are, there are some very easy mistakes to make.

First: don't blindly copy-paste.

Just because AI generated code doesn't mean we understand it or that it's appropriate for our project.

Second: don't skip testing.

If AI helps us write code faster, that doesn't mean we can remove testing from the process.

In fact, generating code faster makes disciplined testing even more important.

Third: don't ignore security.

Especially in WordPress, we need to be careful about things like user input, permissions, authentication, data handling, and anything that touches the database or external services.

Fourth: never share API keys or secrets with AI.

If your prompt contains credentials, tokens, passwords, or private information, stop and think about where that information is going.

And finally:

Don't assume AI is always correct.

It can sound extremely confident and still be wrong.

That's probably one of the most dangerous characteristics of AI-assisted development.

It doesn't always say:

“I don't know.”

Sometimes it gives you a very convincing answer that you still need to verify.

So the developer remains the person responsible for the final decision.

“And if you remember nothing else from this talk, I want you to remember these few things.”

26 of 27

Key Takeaways

AI removes repetitive work—not critical thinking.
Strong prompts produce stronger results.
Repetitive instructions can be reduced through skills.md file.
Review and test every AI suggestion.
Free tools are enough to get started.
Open-source AI offers privacy-friendly alternatives.
Choose tools that match your workflow and budget.

So let me bring everything together.

The first thing I want you to take away is:

AI removes repetitive work — not critical thinking.

That's probably the central message of this entire presentation.

Second, better prompts produce better results.

But don't stop there.

Better context produces better results too.

That's where something like skills.md becomes useful.

Third, review and test everything important that AI produces.

Treat AI-generated code with the same engineering discipline you'd apply to code written by someone else.

And you don't need to start with the most expensive tools.

Free tools are enough to start experimenting.

And if privacy or local execution is important to your workflow, open-source AI gives you another option.

Finally, don't choose an AI tool because everyone on social media is talking about it.

Choose it because it solves a problem in your workflow.

For me, the goal isn't to become a developer who writes code entirely through AI.

The goal is to become a developer who can use AI effectively without giving up engineering judgment.

And that's the mindset I hope you take away today.

27 of 27

THANK YOU

Find me digitally:

https://swarnikashakya.com.np/