Has anyone tried using Claude Code / Ralph Wiggum to package Cloudron apps?

[jdaviescoates]

It sounds like if you get good at very clearly defining what you need done it's pretty good at just going off and doing it:

See for example this write-up which has prompted this thread:
https://www.second-breakfast.co/blog/ralph-is-eating-the-world

@charlesnw had previously shared a CloudronPackagePrompt.md he'd be using but that seems to have moved, but @canadaduane posted a copy of a version of it here: https://forum.cloudron.io/post/108425

I'd guess that @timconsidine has been harnessing AI to do all the packaging he's being doing recently too and my have some insights to share?

And @marcusquinn seems to have lots of AI agents doing his DevOps as per his https://aidevops.sh/

[timconsidine]

@jdaviescoates very interesting links, thank you.
Need to read them again to fully digest.

And yes, I use AI to assist packaging !
Due to illness, I've taken a break from my main project that will make me rich (ha ha ha), and so have had time to do some Cloudron packaging. AI has certainly helped.

My experience is mixed, but increasingly very positive. Some of it aligns with the article you posted.

journey

• I tried AI as standalone resource which helped with individual tasks, but slow and inefficient, and no leverage obtained through building knowledge.
• then I discovered Windsurf as an IDE with integrated AI engines, and big gains from leveraging the conversation flow
• I fell out of love with Windsurf just prior to their acquisition because of coding mistakes and flawed implementations
• often this was actually when using Claude, which many rave about, but my experience has been that it is nothing to get too excited about
• I discovered TRAE and now use this IDE with their AI in auto-mode exclusively for projects, and other standalone engines for ad-hoc tasks, and for sanity checking TRAE, and for planning (mostly Ollama/qwen3-coder:480b-cloud). I forget exactly but I think TRAE has stopped using Claude.

learnings

• AI is magical .... except when it isn't
• detailed specification really does matter, but don't get bogged down, some things can be delegated to its discretion
• initially ban your AI from creating code (!!), ask it to plan the overall task, and then the sub-tasks
• actually, first tell e.g. Ollama what your plan is, and ask it as PlanAI to explain the task for your chosen DevAI, asking DevAI to scaffold overall and sub-tasks
• point out errors, misunderstandings (maybe due to unclear specification) and issue reminders (e.g. Cloudron RO/RW constraints), maybe asking for a revised plan, maybe asking DevAI to consider different approaches to arrive at a best approach
• then instruct DevAI to start coding to implement the agreed plan
• watch it like a hawk to keep it to the plan
• this methodology plays to my strengths (I'm not a pro-dev) and AI's strengths (getting hands dirty)
• intriguingly TRAE does seem to learn cross-project, I can be pleasantly surprised that in a later project, it seemed to learn from mistakes in a previous unrelated project

The whole AI as developer landscape is changing so fast. I am probably missing some tricks and new tools, and I struggle to keep up.

I don't currently buy into the "go to bed and wake up to a finished product" but maybe it will come ... or maybe I am missing something.

[jdaviescoates]

@timconsidine said in Has anyone tried using Claude Code / Ralph Wiggum to package Cloudron apps?:

initially ban your AI from creating code (!!), ask it to plan the overall task, and then the sub-tasks

This is part of the work flow outlined here too:

[marcusquinn]

The coding isn't really the hard part any more. The testing and decision-making still remains a human in the loop domain.

We're going in the right direction, though.

IMO it won't replace developers, it'll just mean developers move up the food chain for creating and managing value by guiding and overseeing coding and devops agents.

Testing and gotchya checking still remains reliant on human experience.

Plus, like flying, you always want to know there's a pilot with a stake in survival.

[robi]

This works because you're dealing with a pile of low quality training data.

This is why we use tons of water to flush sewage/dirt/etc until something we want (diamond/gold) is left over.

Subtractive process like a sculptor removing debris to reveal the statue within the rock.

Requires lots of filtering. Just like with Ralph Wiggum

[marcusquinn]

Reminder to use the context7 mcp in your vibe coding, as that will have the knowledge of all the latest cloudron documentation.

[robi]

@marcusquinn I wonder if @timconsidine uses it.

[timconsidine]

@robi / @marcusquinn I haven't been using context7 MCP but I probably should - not been using any MCP, but again I should - will add to my list - just tell my TRAE to go check docs.cloudron.io/packaging

[jdaviescoates]

@timconsidine said in Has anyone tried using Claude Code / Ralph Wiggum to package Cloudron apps?:

TRAE

https://www.trae.ai/ I presume?

[timconsidine]

@jdaviescoates yes, that's the one
there are others (previously used Windsurf)
But I like TRAE

[scooke]

@timconsidine That packaging link gives me a 404.... Ah, wait a minutes, context7 would just use the text that's under that link, correct? Like /addons, /cheat-sheet, /cli, /manifest, /publishing, /tutorial... and you just append the same command, use context7 at the end of the prompt in the Solo prompt box? I assume your prompt includes Cloudron somewhere. How verbose is it? Like, "Make an app following https://docs.cloudron.io/packaging/ for bookwyrm"? This stuff is fascinating.

[timconsidine]

@scooke I haven't got round to using context7 yet.

The packaging link was typed off the top of my head.
I think the main problem is that there is not a "root" page for the packaging docs. IMO Cloudron should create one.

Let me investigate a bit more later (fighting Zoneminder packaging at the moment)

[timconsidine]

@scooke I checked out context7 - looks good !
thank you @marcusquinn for bringing it to all our attention

easy sign up and a free plan for 1,000 requests per month (don't know how quickly that gets consumed in practice)

@scooke, it seems someone already ingested docs.cloudron.io into context7 (@loudlemur raised this previously, although I missed that post)

so there's effectively no setup to do. TRAE supports context7 from its Marketplace (no cost) and it seems context7 supports all the major coding assistants (if TRAE is not your thing).

I am using TRAE Pro plan, not the SOLO mode - thinking about subscribing for SOLO but currently the Pro plan Builder (now Builder with MCP option) is rocking it for me.

and yes, I'm guessing that a simple prompt in TRAE AI chat after adding the context7 MCP might work :
"use context7 to review cloudron docs and help me package bookwyrm (GitHub : github.com/<dev>/<app>) for deployment on Cloudron"

I will need to check it out further.
After my success with packaging 2 complex apps Zoneminder and Seafile, I'm keen to see what other "impossible" Wishlist items can be ticked off.

EDIT : and it would be wise to add a rule in your AI assistant to use context7 without asking :

Always use Context7 MCP when I need library/API documentation, code generation, setup or configuration steps without me having to explicitly ask.

[marcusquinn]

@timconsidine I'm guessing TRAE is using Sonnet to make it so cheap?

Have you tried Opus 4.5? I'm hooked on opencode.ai + Claude Max.

[robi]

This just in: The Full Nelson - Beyond Ralph
https://github.com/jusgou/nelson

[timconsidine]

@marcusquinn to be honest, I am not sure what model is used because I use it on Auto. Allegedly that means they use the best for the task.

I think they don't use Anthropic, but you can configure access by your own subscription.

Looking into opencode.ai today

[marcusquinn]

@timconsidine cool. 5.2 Codex has been the only ones from those i can trust with coding, almost as good as Opus 4.5, just slower.

I think you'll like. You might like my https://aidevops.sh to give it a full devops lifecycle for coding with proper git and worktree usage. Must saver for parallel work on the same repo, and then connect up all the code auditing tools on your repo for a postflight check and refinement.

[timconsidine]

Looking into opencode.ai today

opencode was a lot easier to set up than Claude Code
And it looks nicer
Thank you @marcusquinn

Looking at aidevops ...

[marcusquinn]

@timconsidine open-source wins again

[nebulon]

I can second this, opencode was pretty straightforwards to get to work against an ollama running on a different machine locally. However my test system is absolutely too old to make proper use of it. At least I know it is theoretically possible if I would get a proper setup