Ask HN: What is the AI setup for an experienced dev starting on a new project?

[postexitus]

I have been working as a software developer for more than 20 years now. Been around the block a few times. Thanks to generous allowances in my current place, I've dipped my toes into LLM world and am quite impressed with it. However, we are on a legacy code base, LLM impact is limited in a sense. If I was to start a new project (let's say a web app with backend) - what is the state of the art for the setup? From Issue tracking, CI, automated deployments, agents, documentation etc - is there some blueprint I can just deploy and know that I am standing on the shoulders of giants?

[mattbrewsbytes]

Your job as an experienced developer is to ship working code, the fact that AI tools are here doesn't change that. I would strive to create as minimal a set of artifacts for AI as possible. Observing where we are today with 6 months or 12 months ago means the state of the art is constantly shifting. Spending a lot of effort implementing some concept you've read about or watched others talk about online might not be useful in 6 months. Its hard to discern valuable content vs. stuff people are finding gets lots of clicks these days.

If what you're creating isn't shippable to a customer/user then really scrutinize it. Things like AI skills, AI task loops, etc. those are all throw-away artifacts, the equivalent of Jira tickets. If you can get AI to help delivery without them, do that. The AI lab companies want you to use tokens but if you can automate something without AI in the mix, since you're experienced I assume you know shell scripting, automation concepts, etc., then automate it to work without token use. This is especially if you are doing something solo - your process and workflow can change on a whim so the least amount of overhead is going to keep you lean and focused on shipping software.

[postexitus]

True, but I see how much of a leverage AI tools give - I would like to exploit them as much as possible and only use my experience where necessary (not necessarily when typing code). I can individually instruct Codex or agents to do somewhat complex work and review and merge, but I am wondering if there is a better way to set everything up.

[nimblist]

For me right now it's this but I'm constantly learning new approaches/improvements

Claude Code running on a dedicated debian lxc with tmux so I can pick sessions up from any machine. The key thing for me is to ensure that claude is constantly maintaining its own project documentation so that it's not reliant on any one session context. I've also set up custom commands to set up a clearly defined plan md file and to incrementally step through it updating completion as it goes.

From my perspective it's important to treat it as another potentially fallible team member. I have extensive unit tests, automation tests and SonarQube carries out analysis on all code merged to main.

Finally github actions to provide a proper ALMS with a non production environment for e2e testing.

[warrenshi]

The thing that changed the most for me wasn't the coding setup.

It was the order.

A couple years ago I'd think:

architecture database implementation deployment

Now I spend much more time before that asking:

does this map to a workflow people actually repeat?

AI made implementation so much cheaper that I've become much more cautious about building the wrong thing quickly.

[CobaltFire]

Very interested in the answers, as I'm in a similar situation.

Cofounder is using agents, I'm directly interacting with Claude and Codex. Self hosted forgejo due to Github issues.

[postexitus]

What is your CI like? Do you auto deploy somewhere?

[CobaltFire]

This is where I show my ignorance.

I'm coming to modern software development late. My experience was in things which were in the mini-computer and COBOL genre of machines, or on things that were so tightly controlled that the review process for one change could take a year (combat aircraft and drones).

So CI/CD is something I'm actually trying to grasp because it feels so alien to me.

All that to say: Don't currently have any.

[tstrimple]

Don't make the mistake of thinking CI/CD is more complicated than it is. It's just automated compilation & testing plus automated deployment. The jargon can be overwhelming. You can do this with forgejo actions. I'm also using self-hosted forgejo and I generated a token for CC to use to manage and setup everything for me.

When code is checked in, it tries to compile it to make sure there are no fundamental issues building the software. This, importantly, happens on not your computer. This helps protect against "it works on my machine" but fails for someone else. If you've got tests, and you should have tests, it will run them for you to make sure those complete as well. If both of those pass, it reaches the CD portion of the process and typically just zips up the working build and does whatever is needed for a deployment in your environment automatically. There are tons of other things you could do as well, like run a linting tool to make sure the code is all formatted consistently across the project or run security checks to make sure passwords or .env files aren't included by accident. But the core concept is build the project, test the project (with various different tools), if it passes deploy the project to some environment where you can see the changes. This drastically cuts down the feedback loop time and increases how fast you can iterate on changes.

For my setup, I've got forgejo hosted on my NAS along with various web apps as containers in podman which I use cloudflared tunnels to selectively make them public. I've got a Claude skill which knows about my hosting environment and can setup / manage new web projects for me. When I add a new web project, CC runs a script that CC built to configure the build, test and deploy actions. After that, any pull request to that new repo will automatically build and test and deploy automatically to a new subdomain. CC could call the APIs directly and handle everything without scripts, but that's where some of the non-deterministic aspects of LLMs might creep in. I want it to setup everything the same way every time so all the apps are consistent. So it uses pre-built scripts for those aspects with the direct API calls being an escape hatch if needed.

[CobaltFire]

Thanks for that. Thats what I thought, but havent implemented.

[runjake]

My sweet spot is the Codex app running separately from whichever IDE I'm using (VSCode or nvim). Codex is pointed at whatever work folder I'm working on.