Fowler: DSLs are not dead, they are what makes LLM usable

In plain terms

Martin Fowler's thesis, published on July 15, 2026, is simple: LLMs produce non-deterministic and sometimes incorrect code. A general language tolerates this drift and propagates it. A specific language (DSL), on the other hand, bounds the space of possible outputs, makes validation trivial, and allows LLMs to generate without having to review each line to pray.

Why it matters now

Two trends are intersecting:

1. LLMs are good enough to generate on a constrained vocabulary (grammar, JSON schema, EBNF) with a very low syntax error rate.
2. The post-2023 software engineering community has largely abandoned DSLs, deemed too costly to design and maintain.

The game has changed. The cost of designing the DSL remains the same, but the benefit has exploded - because a well-designed DSL transforms "having an LLM write code" into "having an LLM fill out a structured form". The difference in reliability is of an order of magnitude.

Under the hood

• A DSL is more useful than a JSON schema when the logic is procedural (if/then, bounded loops) - because it captures business invariants in its grammar itself.
• The most profitable implementation pattern: LLM → DSL output → interpreter or compilation to action. You never leave your strict programming language - the DSL is the intermediary.
• Cases that don't work: creative tasks where the output universe is not enumerable (generating marketing text, brainstorming a name).

So what

For the architect: reconsider the decisions "we'll parse free text from the LLM and hope". Instead, do: LLM → DSL → runtime. Regressions go from silent to glaring.

For the team building an LLM product: the next engineering effort is no longer in prompt engineering - it's in designing a mini-language that captures what your domain allows. It's a return of the language designer profession, under a less prestigious name.

For long-term debt: a well-thought-out DSL survives the next generation of models. A giant prompt, no.