Vibe-coding in OCaml: a bleg from a retiree

Claude Code is $20/mo which is sufficient to get you started. You may hit usage quotas and start paying for extra credits on occasion which is fine for incremental work. The $100/mo plan is more than enough for big new projects where the agent writes all the code in one shot. It’s not needed for smaller tasks. So it’s not free but still cheap in the hands of an advanced programmer, not a vibe coder.

My advice is to start using it on a feature or a project that you wouldn’t create otherwise because it’s too big or not at the core of your interests. There’s not much to lose in trying.

Like everyone else, I’m still discovering how to work with AI coding agents and I expect that things will keep evolving. With that in mind, here are some patterns that have emerged:

  • writing designs ahead of time before asking the coding agent to get started saves time. Make sure the agent has a clear sense of what to do using proven techniques before it starts. Do some research beforehand by chatting with your favorite chat agent to come up with an architecture or a plan that’s detailed enough to be implemented using common idioms. Then feed the plan as a prompt or as a design document to the coding agent.
  • getting from zero to a prototype that works well enough for a demo is great. If you cut too many corners, you may want to throw away much of that code later and it’s ok. Be upfront with your peers that it’s not a final implementation if you want to reduce the amount of pushback. Or don’t show them the code at all.
  • application code that nobody’s code depends on is more tolerant to sloppy designs than libraries. Especially if the application is for personal use, go for it. Make that little utility you’ve always wanted for yourself.
  • implementing a definitive spec or a reference implementation with tests works great. You’ll have to focus your energy on refining the API or user interface but the implementation will mostly write itself.

On the social aspects: there’s a lot of anxiety. Some people are extremely enthusiastic. A few are having a very hard time and react in extreme ways. Personally, I was extremely enthusiastic when I started. It’s cooled down a bit since because the benefits really depend on the type of project.

On team structure: the human programmer is now an architect and the coding agent is another team member. Having another human review all code is not sustainable and kills most benefits of AI coding. Junior programmers should probably still have all their code reviewed so it may be best if they use AI coding sparingly.

Seems like Jane Street is on the same Formal Verification track here:

Get on with it Chet :slight_smile:. Just poke put those money - at least for a month. I think all you need is one day to realize the potential. It is not only that AI providers that generate the hype. It is the users, I have personally not written a line of code since November 2025. It is all prompting from here on. I do like to code …, but when I can do something more easily I am drawn to it. The fun factor changes: from coding to the effect you can cause with little work. Like instead of typing “match exp with ,…” you type: “write me a grammar, an AST, and a parser for this pseudo code grammar sketch”

That was missing from my CS curriculum.