Structured Prompting (narphorium.com)

The piece argues we’re entering a “second software crisis” driven by large language models and exploding prompt complexity, and that we should revive the lessons of structured programming to build new abstractions for natural-language programming. It maps punch-card-era practices to tokens and context windows, showing how ad-hoc long prompts evolved into tool calling (like system calls), the Model Context Protocol (MCP) as an API layer for LLMs, and recent features such as Claude Code’s subagents that provide isolated agent trajectories with local scope, parameters and return values—akin to subroutines and a call stack. Technically, the article highlights three key levers: (1) tool calling and MCP reduce prompt bloat by fetching only relevant data on demand; (2) subagents restore scope isolation and composability so subtasks run with minimal context; and (3) task lists externalize planning so decisions are inspectable and verifiable. Remaining challenges include accumulated agent-history blowup, the need for automatic context compacting/summarization, and formalizing boundaries and interfaces between subagents. For the AI/ML community this signals a shift from prompt tinkering to engineering principled abstractions—scoped subagents, protocolized tool interfaces, and traceable plans—that make LLM-driven systems readable, maintainable and auditable at scale.