I was actually working on something like a "claim checker" recently. Inspired by stuff like Codeql. But I realised that all the really valuable business facing claims can't be checked just by static code analysis, and need to be done via integration and unit tests only. Given that claude code can pretty much write all the runtime tests fairly easily and quickly, I don't see how anything else can be better.

have you actually tried breaking something on purpose and seeing if claude's own tests catch it, or is it more that they've just run fine so far?

Fairly interesting! A few questions:

- When asking the user for an input, does it support multi turn user conversation?

- Who is this made for? Developers would much rather just use python directly with coding agents; non developers would probably still be afraid of seeing the python-ish syntax in the md file + they wont be able to use existing coding agents cause there would be no examples of this new language in any of the LLM's training data.

But it's an awesome attempt! I like the idea overall, and the design of the website!

- When asking the user for an input, does it support multi turn user conversation? Yes, it supports both multi-turn conversations from the llm and you can programmatically ask the user things too -> https://www.banyango.com/margarita/latest/user_input/

- Who is this made for? I think this is crux of my issues so far here. I have been targeting developers who maybe don't see as much value in this as you mentioned.

That said I think that there is a decent cost to developers for maintaining code each time you want more determinism, or don't want to spend tokens on repeatable steps. My hope was that having a lower cost solution closer to markdown would be a nice middle ground.

What if my network is not on the app?

For most people, they never get any feedback from these platforms anymore.. there are just SO MANY PH and HN launches everyday.. that unless you pay some big PH hunter, or are a part of YC (for HN launch help), your post will just get ignored.

It really depends on your product, but cold emails / DMs still works if you try enough times. Launching on product hunt, HN at least gets some eyeballs and maybe some of those eyeballs convert to people who try your product.

Cold emailing could work. But I wouldn't say it's worth it, because the effort you put into building a mailing list could be spent on something more productive and effective. Really depends on the team size.

Product hunt is debatable. I feel like it's a platform for more established products rather than new ones. Launched once, went crazy on socials, and eventually I got 5 upvotes, which is pretty decent already for a small SaaS.

Does anyone know if it won't leak information across slack channel since it has the memory feature?

I ask it to write unit test with close to 100% coverage. This has worked well for me so far

How do you unit test a UI?

It will always be important as a networking / social experience with very like minded folks. But for learning the subject, not so sure

As models get better and better at following instructions, don't you think this will become lesser and lesser useful?

The better the models, the fewer guardrails they will need. However, the more you know about the process the more useful the guardrails are. Guardrails may become less useful over time, but you pay for it in uncertainty and likely token usage.

How is this better than higgsfield?