Indeed, any language that encourages you to program by building tailored DSLs for every problem domain has the same issue (Lisps, Racket…).
It sounds very productive to always be able to express solutions with the perfect language for that type of problem. But you end up with a myriad bespoke languages that are very impractical to document and learn, and are a constant friction for collaboration (even with your future self).
It can work if you think carefully about module boundaries so that the dsls do not leak out of their containing scope. Most of the time, dsls are not well bounded and even if they are, they are likely to confuse newcomers to the codebase. Of course, whether or not dsls are used, every program does have custom dsls in the form of functions that newcomers also aren't going to understand so it isn't as though programs written in a vanilla host language don't have many of the same comprehensibility problems as those written in a dsl.
That being said, not every piece of code needs to be accessible to everyone. Dsls allow you to create a hermetic world inside of a programming environment in which certain properties always hold, even if the environment does not guarantee these things for you. Programs written in the right dsl can be more productive to write and have much lower overall operational costs. But the price is that the rules of the dsl are not necessarily the same as the rules of the host language and this leads to confusion. For the truly ambitious programmer, it can be worth sacrificing accessibility to the median coder to build systems that the median coder can't even conceive of. Of course your reward is to be derided for not expressing the solution in a language they can already understand.
Wrt to Forth itself, Chuck Moore has pointed out that Forth is a multiplier: it makes good programmers better and bad programmers worse. I personally care more about the former property and frankly don't want to work with bad programmers anyway.
To me it is an issue of documentation mainly, smart programmers cannot read minds either. There are standard and effort-efficient ways to document function APIs, whereas DSLs need full user guides. They have complex rules about what are valid expressions and what they do, hard to guess without a thorough explanation, and incompatible with modern dev tooling that makes it easier to quickly to discover how to use an API without much reading or trial-and-error.
I do agree there is room for such approaches and they can have 10x effects in some situations. But more often than not, smart programmers can get wrapped up in the logical beauty of it all, distracted away from the core problem, feeling extremely productive but barely getting anything real done (like Vim can make feel so much faster but you really aren’t when you measure and compare properly).
Creating new languages is useful: config, markup, query… But it is hard and laborious to do well, and usually it is only successful if creating the language is your core goal, rather as a one-off auxiliary part of solving a problem, as languages like Forth or Lisp constantly encourage by their design. And, again, a non-trivial language without a manual is unusable (or it is trivial).
Limits to flexibility can be very useful: they make you focus on solving the problem with the tools you have, rather than making the perfect tools to solve the problem, tools that you will probably never use again because they are too tailored to that one problem.
You can use tailored DSLs in any language. Not just Lisp.
I do it all the time. It is a huge productivity boost.
An example is a biz application I recently delivered to a large international customer. More than 90% of the C++ and Typescript code was auto generated from a simple declarative spec. The only custom code was the customer specific biz logic and custom interface logic to external tools.
Yes it isn’t generated with macros Lisp style but it works equally well.
It sounds very productive to always be able to express solutions with the perfect language for that type of problem. But you end up with a myriad bespoke languages that are very impractical to document and learn, and are a constant friction for collaboration (even with your future self).