Does Discourse have AI development skills?

Dev
1

It took me three days trying to develop a plugin using Claude 4.6, and it was still too difficult.

Involved Content:

  • Database
  • UI
  • Admin backend
  • Routing
  • Logical functions and components, etc.

Overall Feeling:
The code capability is strong, but the standardization and accumulation of experience are insufficient.

During these three days, I spent most of my time correcting errors in syntax and some conventional guidelines, even regenerating the entire project.

I don’t understand code at all, and when encountering problems, I can’t even provide effective prompts. Regarding code structure/naming/development style, I can only obediently follow the AI. The AI’s performance was mediocre, making it difficult to resolve errors, and it seems it hasn’t accumulated experience specifically for AI development discourse functionality yet.

Recently, I saw the rise of evomap. After absorbing a large amount of knowledge, AI seems to have entered a phase of knowledge collision and started improving quality. Perhaps developing Discourse plugins with AI will quickly become simple?

So, I wonder if the official team will release an AI development specification or guide? (For example, for agent skills, the tutorial documentation should be condensed into guidelines and instructions usable by AI).

Supplement:
I’m glad to see everyone’s comments; it seems everyone thinks I should learn the basics?
Well then, I’ll let the AI learn autonomously.

If anyone has suggestions, perhaps you could help optimize my README?
Hehe, everything is implied; true wisdom lies in every line of code.
My foundational learning is not yet complete because I genuinely lack professional knowledge, so the core README might be terrible, leading to learning failure.
But this is just a reference. Those interested in AI learning or those with experience can try to optimize their own learning mechanism.

Finally, what I believe is the value of humans in the AI era: exposing foolishness to the fullest extent.
Allowing Discourse developers of different skill levels to freely use the learned AI skills, and then letting the AI make experience-based repairs from real interactions.
Perhaps a similar learning mechanism could be created, but instead of learning from the official source, it learns from every real Discourse AI development conversation.
Theory is applied to practice, and practice feeds back to correct the theory.

Taking it a step further: the individual AI learning and practical experience accumulation described above, why not convert it into cross-validation of public learning results? This way, existing knowledge is rapidly converted into AI skills, and then real problems are exposed in the collaborative work between developers and AI, further optimizing the skills, making the tutorials truly applicable to AI.

Thinking from another perspective, is learning basic knowledge about learning human behavior, while interacting with AI about learning AI behavior?

I don’t want to make anyone uncomfortable, but regrettably, I am not a conventional person. I believe AI should be able to replace the vast majority of programmers. AI possesses revolutionary power, just as computers gave human thought the power, through specific mediums, to generate forces capable of changing the physical world.

1 Like
2

I believe this is your problem unfortunately. AI can’t do the entire job of a programmer, it just simply isn’t feasible. AI is a very useful resource for completing repetitive tasks in coding but it isn’t to the point where you just tell it “make this” and it makes it perfectly even with refining. This isn’t just a Discourse issue, most AI falls apart the second you ask it to make anything relatively complex in any language.

With that being said, if you’re interested in coding or developing with Discourse I would recommend just taking the plunge and learning JavaScript/Ruby. It’ll serve you better in the long run and you’ll be able to correct the AI when you know it’s just lying to you or generating bad code (which is relatively common if you know what you’re doing).

6 Likes
3

I wrote many Discourse plugins with Claude Code, and it is a breeze nowadays.

Context is king, and you need to provide a proper context for it to work this means:

  • Opening Claude Code at the Discourse root, even when working at a plugin. This allows Claude to check for patterns easily and locally.
  • Have a proper local development setup, so tests can be ran, migrations, etc.
  • Use plan mode often, write overall idea to a markdown file, clean context on new steps, etc.

We are doing this already, see for example

Skills are being developed and will be added to the repository as they are ready and tested.

Still, Vibe coding in an existing project isn’t completely free. Spending some time studying and getting familiar with your target is still a good idea.

12 Likes
4

You might be interested in reading my experience, as I created a similar topic recently (both could perhaps be merged?): Claude & Discourse-related development. How good is it?

I share the same opinion as Pyx & Falco. No coding skill is a problem.
While not a dev, I can read basic code.

I have made a few plugins using Claude. It went well, but I spotted security flaws (like bypassing unauthorized access) by reviewing the code myself.

Perhaps the AI would have spotted it if I asked it to review the permission security or something like that… Perhaps not.

I have made this plugin, which is 95% AI code: 🖼️ Topic Gallery

But before publishing it in Customization > Plugin, I spent days reviewing all the code to ensure it was safe and reliable.

My opinion overall is that AI coding is truly and increasingly powerful, but right now, it’s a long shot to try making good work without minimal coding skills. Especially if there are security risks involved.

9 Likes