Discussion: Project development philosophy, workflow, and AI governance

[ZamenWolk]

Hi Jason,

First off, I want to thank you for taking up this integration and giving it new life. Motorized blinds were the very first IoT devices in my home, and I see a huge amount of potential for the complex logic enabled by this project!

While exploring the repo, I noticed the CLAUDE.md and HANDOFF.md files. It looks like you’ve established a structured "AI-driven, human-orchestrated" workflow. As a fellow developer who is highly interested in the responsible and effective use of generative AI in software engineering, this is the kind of approach I use as well.

I'm opening this issue simply to ask openly about your philosophy and governance policy regarding this project, specifically:

• Workflow Dynamics: How do you balance human architectural steering with the agent's generative execution? Are you mostly acting as the "Product Owner/Architect" while Claude acts as the execution engine?
• Vetting and Oversight: Given the high velocity of commits and the complex solar math involved, how do you handle the review process for generated code? Do you rely mostly on the extensive test suite, or do you have a specific manual vetting philosophy?

To be absolutely clear: I have zero negative bias or hostility here. I am opening this discussion because we've all seen hobby projects where the AI is given free rein, and the maintainer lacks the control or knowledge to validate the generated changes. Looking at the rigorous structure of this repo, I have every reason to believe this is not one of those cases.

However, I think it would be highly beneficial for you to explicitly share your approach. It would give current and future users confidence that this codebase is actively guarded and overseen by someone with a strong grasp of software ownership and development standards.

I’d love to hear your thoughts on how this model is working out for you, whenever you have the free time to share!

Cheers

[jrhubott]

I've been a software developer for over 30 years. Currently I'm a VP of Engineering at a medical device company and don't do much SW development anymore.

I will be honest that I'm leveraging AI for quite a bit of the work. I have a very extensive HA setup but have also struggled with cover automation over the years.

As for the AI. I'm struggling on balancing the "ease" of rapid feature development vs. the maintainability of the code base. Going forward I do plan on being more carefull on the feature expansion. I will though continue to refine the AI use on the project as for me this is also an exploritory use of the technology.

The focus over the next couple of months will be code cleanup and stability. I don't plan on much for feature addition other then exposing more control surface for HA automations to configure the extension. I'm hoping that will reduce the ammount of built in features that need to be maintained.

[ZamenWolk]

Hi,

First, thanks a lot for the transparency and honesty.

It's only fair for me to give a bit more context on myself: I'm a software engineer/architect and business analyst for 7 years or so, in the IT department of a semiconductor company.

Given your response, it seems that you are in the same boat for this project than my colleagues and I are in our work life. I'm far from an expert in leveraging GenAI (yet ?), so I believe there are more expert people on this out there.
My current view is that it's a good thing to leverage AI for the writing itself, even extensively, as long as there is someone behind the wheel that has the experience and "leadership" expertise to know what is appropriate to do and which development processes lead to good code quality.

Given your approach to have extensive automatic tests, the "human-sourced" interventions you provide both in the issues and on the more critical aspects of the codebase, and your reported background in computer science, I have every reason to believe you are approaching the topic right. Especially when using AI, being an expert in whatever language with daily direct experience is not so beneficial, as long as you still remember from your years of software development what is good and bad practice 😉 And for sure, what better than a hobby project to experiment with new things or get acquainted with them.

I look forward to what's upcoming for this repo, and I will keep an eye out for your future projects for sure. I wish I could contribute more to such projects personally, but it's difficult for me to put consistent effort in this activity. So I can only thank you for your service to the community, and that it is much appreciated. 🥂

[jrhubott]

Appreciate the kind words, and the thoughtful framing of the question — it's the kind of thing I'd rather be asked openly than have people wonder about.

Your read on it is pretty much right. The test suite is load-bearing for me: I won't merge anything that breaks it, and I lean on it hard when the generated code touches solar math or the pipeline. For the architectural calls — what to build, how to structure it, when to say no — that's where I stay in the driver's seat.

Hope to see you around as things evolve. Cheers 🥂

[PhilDirty]

Hi Jason, are you operating Claude Mythos or a lighter version?
I have same views here: Need to be careful since ACP has real-life impacts in our homes, but high efficiency for delivery as long as you master it!
Cheers

[jrhubott]

Depends on the change. I have been working on developing skills for distinct operations. For example: Issue-Investigation uses Opus for the investigation, discussion of the identified cause, Haiku for the response to the issue, Opus for the implementation/design plan, Opus again for a Test Driven Development plan, Sonnet for the actual implementation or Opus if the change is identified as Complex, Sonnet to update documentation, and Haiku to cut the release.

Depending on the complexity of the change a secondary code-review is done by Opus. It presents me with it's finding before I allow any correction/change to proceed.

[PhilDirty]

Interesting workflow, thanks for sharing!

[jrhubott]

Quick follow-up to this thread: based on the questions here and on #308, I have added a proper AI-use disclosure to the README and a longer write-up on the wiki at AI-Assisted Development. It documents the skill routing I described above, what stays in human hands, and why CLAUDE.md is intentionally excluded from the repo.

Thanks @ZamenWolk and @PhilDirty for prompting it. The page is better for having the question asked openly.

Note: this comment was drafted by Claude (the AI assistant the new page describes) and reviewed by me before posting.