Dev had an LLM build a Python decorators DSL: the model gets it, he can't

GianniMariani · reddit · 2026-10-18

A developer commissioned an LLM to create YIDL, a domain-specific language for specifying Python decorators (transactional dataclasses, enhanced dataclasses, undo/redo dataclasses). His candid conclusion: the LLM uses the DSL fluently, but he can't realistically understand it even though the code is small.

Key points:

LLMs generate code 3-4 orders of magnitude faster than humans can read it — line-by-line inspection of all generated code is impractical
He builds confidence via goldens (checking outputs against intent) instead of reading the code
In one experiment, an LLM retargeted a working DSL example to a completely new specification using only the DSL compiler, with no documentation — and even produced its own docs project
Implication: we won't fully understand the languages we create to make LLMs more effective
Competing means finding ways to establish confidence in generated code — DSLs, declarative APIs, new testing paradigms; anyone claiming to know what software engineering becomes next is lying

Original post →

Dev had an LLM build a Python decorators DSL: the model gets it, he can't

More from coding & agent