##  [The State of AI Development: From the Perspective of a Seasoned Front-end Developer](/the-state-of-ai-development-from-the-perspective-of-a-seasoned-front-end-developer) 

 On this page

 

This was tough to write, as I don't want to come across as a Luddite or cynical.

I am definitely not a Luddite. And it is believed that some cynicism can be healthy for critical thinking. With the small disclaimer out of the way, let's have a look at the current state of AI.

In the previous issue of my newsletter, I said I want to spend more time doing things I like. I said this is my goal to understand Agentic AI.

## I like Writing (and learn)

The good thing is: I am writing about my learning and building in real-time, and writing is one of the things I like to do. I believe if I can get my work process automated, I can spend more time writing about it; so it's almost a virtuous cycle.

## What is it about writing I like?

I enjoy the exploration process and putting it down in words; I find it creative.

Even though writing about tech may seem mundane, I also feel you can add a little interest to it, make it fun, and share unique ways of looking at and doing things.

In the end, it is all writing, and it is what you make of it.

## Tinker and Learn

Tinkering with new technology is one of the reasons I first got interested in the work, and writing helps to solidify the learning process.

As I have stated before: at some point, you need to start using whatever you are learning, and then things will become familiar and patterns will start to emerge; your brain will start to put things together, make connections, and form an understanding that you can build on.

It's hard to teach that, but once you work it out, you are half way there.

In a way, LLMs are just faster at finding these patterns; that is the whole reason behind the massive scaling that has been done recently. So LLMs should only be seen as an extension of you. Which brings me back to Agentic development.

## Learn First, then play around and Observe, Repeat

Whatever you are doing, you have to learn about how it works first, and there is the catch with using AI.

> The fundamental differences in AI are the models have learned (trained), and now it is pattern matching. It is only doing the pattern matching part.
> 
> Agentic loops, like the Aider coding agent/harness I wrote about, can "decide" if to continue on a task or to mark it complete using tests or other criteria you have given it.
> 
> Tests that either you ask it to create or tests that you have provided. You need to be in control here.

## What you know will influence the LLMs Output

Even though a lot of Drupal stuff could be done in the UI using prompts, using Drupal AI Chatbot/AI Assistant API, I see demos and feel that I could do it quicker by clicking through the steps manually; this is true for almost all workflows when using them to create unique use cases.

Or maybe writing YAML configuration files for Drupal and importing could also be quicker.

The reason being you need familiarity with the system, in this case Drupal, and you also need familiarity with the content or subject matter of the website or application you are building.

In this given example, you need to be very familiar with Drupal and if you use chat it may in fact slow you down, and cost you tokens, but that is another story.

So that's the question you need to ask: Is it quicker to make some elaborate prompt or click through the UI to set up that new field? If you know Drupal I think the latter.

> Can you build me a new page with an area to write my article? It needs to have a good editor as I would like to place an image left and right. I would also like to add aside content at intervals; this content can be other content from the system. It would be good to have an easy way to visualize and add the aside content.

## A Deeper Look at the Mechanics of Setting Up a "Simple" Field

Take, for example, a text field. Should it be plain text or text with a summary? Should it use a maximum length, or should we install the max-length module to limit the length?

> In Drupal, a plain text field will have a max length of 256 chars, you can set a length on creation, before the field has data in it. It is a database restriction for the type of field that is used. Length is set in the UI but this info isn't given to the content creator/manager. The user will only find out when they go to save the page.
> 
> Long text fields can be limitless, and there is way to set a length. You can make the summary a required part of the field.
> 
> If you want to enforce a max-length to any field with a visual indicator, you'll want to use the Max-length module.

## How about Structured Data?

Taking it one step further with structured data, in a Drupal context, can it be a block type, or should it use a Paragraphs type? Or is an SCD for a canvas page the best way?

This is more an architectural decision on whether the component data needs to be reuse-able and/or repeatable. But something you need to decide.

## Next, we have the Content Editing Experience to think about.

With the text field with a summary, once we have it set up:

- Should the manage form widget configuration settings be set to a five-line or one-line default?
- Should the summary be required?

Can you see where I am coming from?

## Prompting Knowledge, Creating usable Recipes, or Setting up an AGENTS.md and SKILLS.md(s)

If I am going to write that much detail in the prompt, maybe I should write a recipe and run an install or import via the CLI and/or drush.

Or do you write a SKILL.md with you hard set preferences?

### Prompting

I shared a demo video a few weeks back, and even though it was impressive, if you know what you are doing in Drupal and know the content matter (which you should), it may have been quicker to do the work by yourself.

**Chat is that, you ask a question, and it gives you an answer, and if you know what you are doing, then you already know the answer, or at least know that the answer given is correct.**

### Agentic Loop Writing

I think the Agentic development might be useful if you have years of documented site-building workflows that build out boilerplate websites for a particular niche or vertical. You know what you are doing.

In theory this is what Drupal recipes were introduced for and is possibly the middle ground.

**Agentic workflow allows an agent autonomy**, but from what I understand, you need to provide skills to an agent to make them useful.

## That's what bought me to search Drupal Agent Skills

That brings me to a search, and you guessed it, Drupal AI agent skills brings up a search.

But hey, the module released on 1 February 2026, is now unsupported.

Moreover, work is continued at [AI Best Practices for Drupal](https://www.drupal.org/project/ai_best_practices). It is extremely early if you read the Roadmap. The project is built around skills, which by power of deduction, this is what best practices are.

[Skill have a specification](https://agentskills.io/specification), yip, the MD file with YAML frontmatter, and a markdown body.

Skills are how humans "communicate" important information to LLMs.

## And, that's where the last newsletter came in.

In my [last newsletter](https://preview.mailerlite.io/preview/786936/emails/188775387250034302), I talked about a basic model for agentic work flows. The model is fairly simple; it is a loop.

The part that will really make your workflows/agentic loops shine will be the skills you provide. The skills are the to-dos and not-to-dos, and once you have these set up, you may be able to get some success.

I think you need to think like I am here. Keep working on refining your workflows manually with the idea they will become skills you can give to an LLM.

Create a skill and once done, use it, test it, clearly document its use, and then contribute to the world.

## You're going to have to work this out yourself

A small note to leave on, I have found success in chat style coding for JavaScript. It is not always able provide a one shot solution, but it does write code much faster than I.

Also, in the last few weeks I have been doing some Drupal 10 to 11 upgrades that have custom modules that needed modernized code. Something I may have once passed on to save time, but now I can do it.

So what I am trying to get across here is not that AI is unuseful; it is very useful.

But it will take a lot to set up Agentic Loops, and sometimes the cost benefit just won't be worth it.

What do you think? Is it worth setting up Agentic workflows? If so, what is the criteria that the project needs to meet?

Alternatively, is this a long haul thing? I think that is it; it is not going to happen overnight. Adding a little bit here and there where it makes sense. Over time, you'll be able to fill in the gaps and build up a fabric; something that can do a big chunk of work unsupervised.

This perspective is from someone who is freelancing; I have limited time to get swept away in the hype and feel I am getting left behind. But hey, some people are either being forced to use AI, or some are new to this game; the vide coder. In the end, I think we will all converge.

It is truly early times.

Thanks for reading

Simon