Don't Parse, Call (timkellogg.me)

The piece argues that freeform prompt formats are dead — stop inventing ad-hoc parsing rules and start using function-calling APIs. Modern LLMs are increasingly "agentic" and are heavily trained to call functions, so exposing capabilities as API functions yields far better reliability, expressiveness, and safety than trying to encode structure in prompt text. Rather than wrestling with brittle instructions (one-line-only answers, banned characters, etc.), you declare functions and let the model invoke them; the model’s calls become a structured, auditable log you can use or ignore. Technically, function calling is now ubiquitous across models even if internal token formats differ; the API is the stable interface. Functions let you capture rationales, represent optional results (sum-type like str | None), model failures, and support multiple or zero calls in a session. That inverted-control pattern is the basis for lightweight agents: you provide primitives and the model orchestrates them. The author also warns against high-level frameworks that hide whether you’re using functions, JSON, or files — prefer low-level APIs (newer ones add document/file concepts and caching to reduce prompt injection and bloat). Bottom line: stop parsing prompt text—design functions and call them.