I’ve been using it to develop firmware in c++. Typically around 10-20 KLOC. Current projects use Sensors, wire protocols, RF systems , swarm networks, that kind of stuff integrated into the firmware.

If you use it correctly, you can get better quality, more maintainable code than 75% of devs will turn in on a PR. The “one weird trick” seems to be to specify, specify, specify. First you use the LLM to help you write a spec (document, if it’s pre existing). Make sure the spec is correct and matches the user story and edge cases. The LLM is good at helping here too. Then break down separations of concerns, APIs, and interfaces. Have it build a dependency graph. After each step, have it reevaluate the entire stack to make sure it is clear, clean, and self consistent.

Every step of this is basically the AI doing the whole thing, just with guidance and feedback.

Once you’ve got the documentation needed to build an actual plan for implementation, have it do that. Each step, you go back as far as relevant to reevaluate. Compare the spec to the implementation plan, close the circle. Then have it write the bones, all the files and interfaces, without actual implementations. Then have it reevaluate the dependency graph and the plan and the file structure together. Then start implementing the plan, building testing jigs along the way.

You just build software the way you used to, but you use the LLM to do most of the work along the way. Every so often, you’ll run into something that doesn’t pass the smell test and you’ll give it a nudge in the right direction.

Think of it as a junior dev that graduated top of every class ever, and types 1000wpm.

Even after all of that, I’m turning out better code, better documentation, and better products, and doing what used to take 2 devs a month, in 3 or 4 days on my own.

On the app development side of our business, the productivity gain also strong. I can’t really speak to code quality there, but I can say we get updates in hours instead of days, and there are less bugs in the implementations. They say the code is better documented and easier to follow , because they’re not under pressure to ship hacky prototype code as if it were production.

On the current project, our team size is 1/2 the size it would have been last year, and we are moving about 4x as fast. What doesn’t seem to scale for us is size. If we doubled our team size I think the gains would be very small compared to the costs. Velocity seems to be throttled more by external factors.

I really don’t understand where people are coming from saying it doesn’t work. I’m not sure if it’s because they haven’t tried a real workflow, or maybe tried it at all, or they are definitely “holding it wrong.” It works. But you still need seasoned engineers to manage it and catch the occasional bad judgment or deviation from the intention.

If you just let it, it will definitely go off the rails and you’ll end up with a twisted mess that no one can debug. But use a system of writing the code incrementally through a specification - evaluation loop as you descend the abstraction from idea to implementation you’ll end up winning.

As a side note, and this is a little strange and I might be wrong because it’s hard to quantify and all vibes, but:

I have the AI keep a journal about its observations and general impressions, sort of the “meta” without the technical details. I frame this to it as a continuation of “awareness “ for new sessions.

I have a short set of “onboarding“ documents that describe the vision, ethos, and goals of the project. I have it read the journal and the onboarding docs at the beginning of each session.

I frame my work with the AI as working with it as a “collaborator” rather than a tool. At the end of the day, I remind it to update its journal of reflections about the days work. It’s total anthropomorphism, obviously, but it seems to inspire “trust” in the relationship, and it really seems to up-level the effort that the AI puts in. It kinda makes sense, LLMs being modelled on human activity.

FWIW, I’m not asserting anything here about the nature of machine intelligence, I’m targeting what seems to create the best result. Eventually we will have to grapple with this I imagine, but that’s not today.

When I have forgotten to warm-start the session, I find that I am rejecting much more of the work. I think this would be worth someone doing an actual study to see if it is real or some kind of irresistible cognitive bias.

I find that the work produced is much less prone to going off the rails or taking shortcuts when I have this in the context, and by reading the journal I get ideas on where and how to do a better job of steering and nudging to get better results. It’s like a review system for my prompting. The onboarding docs seem to help keep the model working towards the big picture? Idk.

This “system” with the journal and onboarding only seems to work with some models. GPT5 for example doesn’t seem to benefit from the journal and sometimes gets into a very creepy vibe. I think it might be optimized for creating some kind of “relationship” with the user.