News for hackers. Stuff that matters.
Re:Ask HN: What's your current AI agentic setup? by bradknowles on Sunday November 08, 2026 @01:51AM (#50411009)
So, at my previous job, I used Cursor+Claude to do virtually all of my development in Ansible. I didn’t understand anything about Ansible, but I was proficient with other tools in this space like Puppet and Chef, and I knew enough about development in general that I could tell the tools how the code should work to do the thing I wanted. And any time it didn’t, I made sure that lesson learned went into the CLAUDE.md file and was applied across the entire repo. And I also automated that rule about lessons learned, so that Claude would automatically update the CLAUDE.md file every time it learned a lesson.

Many lessons were learned along the way, with lots of defensive programming checks to be done, before any operations were actually executed. This includes basic things like bounds checking everywhere, if the variable X was going to be divided by the variable Y, then both X and Y have to be actual numbers, and Y cannot be zero, etc….

I also made sure that Claude would write down everything it had discovered and everything it was planning on doing, into local files, before each and every step. That helped make sure that I would survive a compaction, and Claude could come back and re-read the files afterwards and restart in the same place.

This time, as a volunteer web admin for a non-profit working with Claude Code, I’m also making sure that everything that needs to be done gets recorded first in a TODO.md, and everything that is done gets recorded into a CHANGELOG.md. Everything is done via pull requests, and all work gets committed to the repository on a separate branch while the work is being done. If the remote files get blown away accidentally, then I have local copies. If the local copies get blown away, then I have the repo. The local copies also get backed up to private cloud storage.

I also have all the work done in a given round checked by a sub agent that is free from all of the local context, is given a minimal threat model statement, is told what areas to focus on, and then it is let loose on the new code. Any problems it finds get reported back to the parent and verified there. If verified, that work goes into the next round to be fixed. But we can’t increase the scope of a PR in progress, so if scope would need to be increased then that gets recorded as a new todo. I also have some additional rules that help ensure the PR changes converge, so that the code can actually be pushed out quickly, including a maximum number of rounds that can be performed, and rules about how the number of things that needs to be fixed in each round needs to keep going down.

I’ve already had Claude implement a whole host of security-enhancing GitHub Actions tools, as well as WCAG scanning with `axe`. That way, every push to the repo goes through a whole gauntlet of scans that will block the build, if any problems are found.

I’m trying to take all the lessons I know of that are good for junior programmers to be told to do, and putting them into CLAUDE.md so that I can get Claude to follow them.

I also have an opinionated list of the principles by which I believe that a good website should operate. Like, don’t depend on resources from any external server (e.g., Google Fonts). Use JavaScript only where absolutely necessary. No ads or tracking techniques of any kind. If you link to an external resource that is useful, you should actively identify where that resource is coming from and when you last cached the data from it. None of the standard blocking tools should find anything to block when they load the website. Everything about the website needs to be as accessible as possible. And so on.

I would love to be using more in-depth multi-agent tooling (e.g., Gas City or Openclaw), but I am not yet proficient enough to turn that on. And I am the human gate for actually approving all of these changes, because I do not yet trust the tooling enough.

I’m just getting started with tools like Claude and Claude Code. But I’m feeling pretty good about the work that is getting done, and everything I’ve looked at so far seems to be a solid improvement over what was there before.

Re:Ask HN: What's your current AI agentic setup? by bradknowles on Sunday November 08, 2026 @01:52AM (#50411017)
I am paying for Claude and Claude Code. I do have an old ChatGPT account that I need to shut down. I haven’t used it in a while.