I’m not a developer. I just got basic coding knowledge. I can read simple code. I recently created a plugin, relying mostly on Gemini free. 95% of the code was AI-generated. While the plugin definitely works, and the interface and admin experience are quite alright, I have legitimate concerns about the code cleanliness. Despite my prompting efforts[1] (and manually feeding the AI with excerpts of official code), I’m certain the code doesn’t make great use of the Discourse environment. Helpers, components, classes, and so on. Which is why I occasionally fed it with pieces of official code. I heard that nowadays, many developers use AI in their workflow. I’m curious about how better an AI-generated Discourse plugin code would be if the AI were aware of the Discourse ecosystem. I was advised to try Claude, and more specifically, the Opus model. I was also told it’s very token-consuming and thus expensive. I don’t have real ambitions regarding my projects. It’s just a hobby. I’m curious about how a plugin I imagine would look in the real world, and AIs are fairly good for this. Have you tried using AI when developing plugins or components? Did you try Claude Opus integrated into your IDE? How helpful was it? Does it make proper use of the codebase to generate new code? How expensive is it? Which plan did you go for? writing “prompting efforts” makes me feel awkwardly embarrassed :lolsob: ↩︎

[image] Canapin: I’m curious about how better an AI-generated Discourse plugin code would be if the AI were aware of the Discourse ecosystem. I’ve found that it’s significantly better to work with AI on Discourse if it has examples to work from, and Claude Code is specifically good at this (Google Gemini’s CLI too!). For a hobbyist, the $20/mo plan could probably take you pretty far… it’s not too hard to run into daily limits, but at that point you can either just wait until the limit resets or buy some credits.

I think I was mistaken about the price. The free/pro/max plans are for using Gemini through their web interface. (https://claude.com/pricing). The API price is token-based (https://claude.com/pricing#api); there are no monthly subscriptions for using their API, right? If using AI in, let’s say, VSCode, needs to use their API. edit : well I’ll read some tutorials before writing next time :upside_down_face:

This past week I spent some time trying to write a Discourse plugin using ask.discourse. I was amazed how I could describe the function of the plugin I wished to create and it spit out lots of tips and code snippets, some of which actually ran. I’m a hobbyist programmer. Along side my Discourse instance I run another server with a MySQL database, a PHP backend and a Jquery/Javascript frontend but I am not a proffesional programer. Most the time when developing that site, I just feed plan old google what I want and it (Gemini I guess?) spits out code samples. Most run right out of the box and I am good enough at Javascript to figure why if they don’t. I’ve used many programming languages over the years and in years past I spent hours reading, searching and experimenting to figure out thing that I now feed into google and finish in minutes what would take hours or days in the past. I was amazed how well ask.discourse worked and eventually I got a plugin working and roughly doing what I wanted. With some styling (css) it could/may become a useable plugin. I’m intrigued with the possibility that an AI assisted service could streamline code creation

I wrote a plugin for my Discourse page with Antigravity that displays the most popular anime series. I also created a watchlist. Sometimes it glitches and you have to make a lot of corrections, but I’m reasonably satisfied.

I got it to work on VSCode with a Pro plan (20$). I find the consumption fair. 3h of prompts got me to this: [image] So far, results were not great. :thinking: What could be interesting is describing what should be configured, or what should be told to Claude when creating a plugin or theme component for Discourse (from the skeleton). What instructions should be written in CLAUDE.md to maximize its efficiency? Should certain Discourse directories be prioritized when AI is looking for information in the codebase (I think of controllers, models, serializers, services…)? How to make it aware of the standard files hierarchy, naming, and conventions?

We should ship skills, @j.jaffeux and I think they can be mighty helpful

[image] sam: skills I take it you are referring to [image] Agent Skills Overview - Agent Skills A standardized way to give AI agents new capabilities and expertise. For those new to skills you might find the following of value. Anthropic Claude skills OpenAI Codex skills Google Gemini skills The list is not inclusive and is growing faster and longer every day.

I looked at https://github.com/discourse/discourse/blob/main/AI-AGENTS.md. I have questions! NEVER store find() results - causes stale element references after re-renders ALWAYS lint any changes you make Do AIs understand the all-uppercase emphasis? Would it act differently if written “Never” and “Always” instead? Are things like “architect mode” or other “xxx mode” prompts becoming soft standards in AI development? Do they have any real effect on model behavior? Are they pure conventions? Don’t write obvious tests I can see what an obvious test is, but do AIs understand what an obvious test (or obvious anything) is? I toyed for a couple of days with Claude in VSCode. It’s quite impressive to see everything working by itself. Reading, creating, modifying files, running bash commands, etc… :exploding_head: About Claude’s work on plugin creation, this is what I observed: When iterating over the same pieces of code and solving issues, it tends to use excessively specific variable names. For example, it would name a variable original_url instead of a very fine and simple url, as if meant to point out changes that were made, when it’s not necessary. Iterations often lead to convoluted code, and sometimes, unnecessarily complicated queries. Asking the AI to review and point out parts of code that can be refactored from time to time is proven useful for this :slight_smile: When asked to find solutions for a given issue, I’ve been pleased by the answers. The suggestions seem accurate. When it proposes several options, it can accurately weigh the pros and cons of each one. In SCSS, it (rarely) uses hardcoded colors where it’s not advised, instead of using Discourse’s color variables. It doesn’t seem to be aware of Designing for Different Devices (Viewport Size, Touch/Hover, etc.) and will rely on standard media queries instead. After several tasks have been made, I like to ask the AI to review the code and see if some parts can be optimized without impacting the maintainability. I didn’t have many issues with the results. Sometimes, it tweaks too much and breaks the code. Occasionally, it seems to create unnecessary conditions for the sake of security and robustness, but for situations I’m not sure exist in real-world conditions. For example, in my search-related plugin, it would check if a post has any topics associated with it. But posts without a topic are not supposed to exist in Discourse, I think. If you have one, it means something went very wrong in your instance at some time, right? I feel it made no sense to make this check in my plugin. :thinking: It seems good at scaffolding tests! So, good experience so far, even without using advanced AI configuration. I spent my weekly quota in four days, which is great because it forces me to take a three-day pause :raising_hands: Curious to see how it will change when Discourse’s skills are added.

I spent the day finding my bearings with Claude :heart_eyes: (divorced from ChatGPT) and am looking forward to having a stab at playing around with Claude Code to at the minimum get some prototyping done or scratch my Discourse itches! If there are skills lying in wait somewhere I’ll be happy to try using them.

[image] stephtara: If there are skills lying in wait somewhere I’ll be happy to try using them. Incoming: [image]Discourse有AI开发的skills吗？ Skills are being developed and will be added to the repository as they are ready and tested. After experimenting with plugin developments, I’ve tried creating themes with Claude. The results were catastrophic, but I wasn’t really surprised as it often output nonsense with even simple CSS requests. It doesn’t seem to grasp the context in which each element lies, which results in poorly made or positioned elements, and often a bunch of useless attributes. Until there are specific directives for theme development, I find it very unreliable for those tasks.

Claude & Discourse-related development. How good is it?

Development

EricGT (EricGT) January 31, 2026, 8:45am 8

I take it you are referring to

For those new to skills you might find the following of value.

Anthropic Claude skills
OpenAI Codex skills
Google Gemini skills

The list is not inclusive and is growing faster and longer every day.

Topic		Replies	Views
Discourse有AI开发的skills吗？ Development	5	476	April 6, 2026
How hard would it be for non programmers to be able to use the Discourse AI - AI bot to help them create plugins and/or themes Development ai , ai-bot	10	1000	September 1, 2023
How are you using AI/LLMs to create themes/components/plugins? Development ai	3	299	April 18, 2025
Is there any AI at the core of standard Discourse? Product Support	15	1917	May 31, 2023
Discourse AI plugin with self hosted discourse site Product Support ai	2	583	July 9, 2024

Claude & Discourse-related development. How good is it?

Related topics