The developers who get the most out of agents tend to be unusually good writers. Not in the literary sense — in the functional sense. They write clearly, precisely, and with awareness of how their words will be interpreted by someone who doesn't share their context. These are exactly the skills that make prompts work.

This connection isn't coincidental. A prompt is a piece of writing with a specific reader — the model — and a specific goal — producing a particular output. The same properties that make any functional writing effective apply here: clear structure, precise word choice, explicit statement of what matters, anticipation of misreading. The gap between a prompt that works and one that doesn't is often a gap in writing quality, not a gap in technical sophistication.

The implication is that improving as a prompt engineer and improving as a writer are the same project. When a prompt fails, asking "why did the agent misunderstand this?" and asking "why wasn't this clear?" are the same question. The answer usually involves an assumption the writer made that wasn't communicated, an ambiguity the writer didn't notice, a priority the writer thought was obvious that wasn't.

This also means that the editing discipline transfers. Good writers revise. They read their own work skeptically, looking for places where the meaning isn't as clear as it felt when writing. They cut what doesn't earn its place. They reorder for emphasis. All of this applies directly to prompt revision. A prompt that isn't working usually needs editing, not just appending — more words rarely solve the problem that fewer, better words would fix.

There's a practical exercise worth doing: take a prompt that isn't producing the results you want and edit it as if it were prose. Remove the redundant instructions. Clarify the ambiguous ones. Make the structure explicit. Prioritize the most important constraint. Often the improved prompt outperforms the original significantly, and the improvement came from writing craft, not prompt engineering technique.

The best tool for working with language models is facility with language. Develop it deliberately.