What Kind of Programming Is Natural Language Programming? (dsyme.net)

Natural Language Programming (NLP, here meaning programming by prose prompts) should be thought of as constraint or orchestration programming rather than a replacement for precise imperative/functional code. Well‑written natural language shapes goals, bounds and preferred behaviors without enumerating exact control flow — it supplies useful generality, not formal precision. That makes it ideal for high‑ambiguity, high‑context decision problems (the “Concierge Robot” example: “get me to the airport on time”) where an LLM + tools can ingest calendar, traffic, booking APIs and human messages and make decisions on demand. Precision‑purists (Dijkstra’s Ghost) worry about loss of verifiability, but that critique confuses the wrong use case: natural language’s ambiguity is often the feature, not the bug. Practically, NLP is most valuable in three domains: agentic orchestration workflows (e.g., GitHub Agentic Workflows for CI/QA automation), guided generation (semi‑ephemeral specs or PRDs used to synthesize boilerplate code), and short‑lived task programming (ad hoc “vibe” coding). Key technical implications: systems must be agent architectures with tool access, runtime decision logic, and on‑demand interpretation rather than compile‑time concretization; some ambiguities should be resolved upfront (budget, safety constraints) while others are intentionally left general. Costs include latency, reliability and tooling complexity; limits include domains requiring provable correctness or tight performance. Teaching should emphasize when constraint‑style NLP is appropriate, how to encode essential constraints, and when to fall back to traditional precise programming.