Michael Truong, a developer and author, found that providing a single, canonical writing example to his AI agent produced higher-quality content than an extensive set of manual stylistic rules. While building an AI-assisted content pipeline for technical articles, Truong initially attempted to define the desired voice through a Cursor rule file containing specific tone guidance, pacing instructions, and negative constraints. He assumed that a detailed, encyclopedia-style list of requirements would lead to better results. However, this approach resulted in a failure loop where each revision caused new issues, such as the output sounding like generic engineering documentation or losing all personality as the model over-corrected to follow the growing list of instructions. The rule file reached a peak of 69 lines without achieving the desired tone.

Truong discovered that replacing the 69-line checklist with a link to a single previously published article, titled "Schema first, prompt second: valid JSON wasn't enough," significantly improved performance. By instructing the model to "read the example, match the example," the AI better mimicked structural choices, pacing, and the balance of concrete observations that are difficult to encode in prose. The final rule file was reduced to 23 lines. Truong notes that examples carry implicit decisions regarding context and detail that are often lost when translated into rule-based checkpoints.

This method requires treating the exemplar article as code that may need future refactoring, as using a single example limits the output to that specific format. Truong advises authors to use this approach by shipping one quality article to act as a spec, pointing AI agents to that file, and reserving traditional prompt rules for operational tasks like file management or publishing steps. The author concludes that while prompting remains essential for facts and structure, providing a high-quality example is a more effective way to capture nuances of tone in long-form writing than maintaining a sprawling, contradictory style guide.

Michael Truong, a developer and author, found that providing a single, canonical writing example to his AI agent produced higher-quality content than an extensive set of manual stylistic rules. While building an AI-assisted content pipeline for technical articles, Truong initially attempted to define the desired voice through a Cursor rule file containing specific tone guidance, pacing instructions, and negative constraints. He assumed that a detailed, encyclopedia-style list of requirements would lead to better results. However, this approach resulted in a failure loop where each revision caused new issues, such as the output sounding like generic engineering documentation or losing all personality as the model over-corrected to follow the growing list of instructions. The rule file reached a peak of 69 lines without achieving the desired tone.

Truong discovered that replacing the 69-line checklist with a link to a single previously published article, titled "Schema first, prompt second: valid JSON wasn't enough," significantly improved performance. By instructing the model to "read the example, match the example," the AI better mimicked structural choices, pacing, and the balance of concrete observations that are difficult to encode in prose. The final rule file was reduced to 23 lines. Truong notes that examples carry implicit decisions regarding context and detail that are often lost when translated into rule-based checkpoints.

This method requires treating the exemplar article as code that may need future refactoring, as using a single example limits the output to that specific format. Truong advises authors to use this approach by shipping one quality article to act as a spec, pointing AI agents to that file, and reserving traditional prompt rules for operational tasks like file management or publishing steps. The author concludes that while prompting remains essential for facts and structure, providing a high-quality example is a more effective way to capture nuances of tone in long-form writing than maintaining a sprawling, contradictory style guide.