I built a programming language using Claude Code
A programmer describes creating an entire toy programming language by having Claude Code generate all the code, prompting debate over what it means to “build” software when humans increasingly act as architects rather than line-by-line coders. Commenters argue about whether new languages still matter in an era when LLMs can both write and learn them on the fly, touching on issues like type safety, training data, and whether AI-friendly languages should differ from human-friendly ones. Many raise concerns about the reliability and ownership of fully AI-generated code, as well as the slot‑machine‑like addictiveness of “just one more prompt” development workflows.
Role of Programming Languages When LLMs Write Code
- Debate over whether language choice still matters if humans neither read nor write most code.
- Many argue it does: performance (e.g., Rust vs Python), safety guarantees, and constraints still shape system behavior.
- Others note that if the human can’t read or reason about the generated code, the benefits of a sophisticated language may be lost.
Language Design for AI vs Humans
- Several see more value in new, specialized languages: LLMs can learn niche syntaxes from docs and examples without the usual human-adoption barrier.
- Others counter that without substantial training data, LLMs perform worse, and stuffing language specs into context is inefficient.
- There’s interest in languages that:
- Make invalid states unrepresentable.
- Emphasize concurrency safety and performance “knobs.”
- Are concise but still human-readable (not extreme code-golf; not Java-level verbosity).
- Disagreement over whether terse syntaxes or token-efficient designs actually help models much.
Quality, Testing, and Guardrails
- Strong skepticism about relying on LLM-written code plus LLM-written tests as “guardrails”; tests can be wrong and still all pass.
- Several report that LLMs often hallucinate APIs, mishandle edge cases (e.g., float lexing), or quietly add fallbacks/mock paths that mask failures.
- Formal methods and stronger static guarantees are mentioned as missing but desirable.
Practical Experiences & Productivity
- Multiple accounts of using Claude/Codex to:
- Build toy languages, interpreters, or DSLs quickly.
- Prototype games, frameworks, and large systems far faster than solo coding.
- Others question claims of massive productivity gains, citing studies showing more modest boosts and personal experience of frequent errors.
Ownership, Copyright, and Ethics
- One subthread notes that fully machine-generated code may not be copyrightable under current US guidance.
- Some worry about AI dependence leading to fewer new foundational tools and a potential “long dark teatime” for human engineering.
Behavioral Concerns
- Several compare LLM prompting to gambling: unpredictable results, “just one more prompt” compulsion, and the sense that “the house always wins.”