Subagent Instruction Operationalization

Overview

Translate a completed style specification into a structured LLM subagent prompt that can reproduce a target voice. The core principle: every constraint in the style specification must become either a concrete, measurable prompt directive or be explicitly documented as aspirational. A style specification is an analytical artifact (what the voice IS); a subagent instruction is a generative artifact (how to PRODUCE the voice). This skill bridges the gap by converting each upstream analysis output into a prompt section with verification criteria.

Why this is hard: LLMs respond to instructions with varying fidelity depending on constraint type. Lexical constraints ("use grade-level 12 vocabulary") are moderately enforceable. Structural constraints ("lead with evidence before claims") are somewhat enforceable. Distributional constraints ("maintain a 40/30/20/10 split of explaining/advising/challenging/asserting") are weakly enforceable without few-shot examples. The prompt must be designed with awareness of these enforcement gradients.

Research foundation: Persona-based prompting (role assignment) produces measurably different output style, tone, and vocabulary (Brim Labs, 2025; PromptHub, 2025). Detailed persona specifications consistently outperform vague ones, and LLM-generated personas often outperform human-written ones (LearnPrompting, 2025). Few-shot examples are often more effective than lengthy instructions for style constraints (Palantir, 2025). Stylometric research confirms that function-word patterns, sentence-length distributions, and punctuation habits are the most stable authorship markers -- and therefore the most important features to replicate (Mosteller & Wallace, 1963; Kestemont, 2014). Recent work on LLM style transfer shows that instruction-tuned models can approximate target styles with one-shot accuracy between 67.6% and 94.7% depending on prompting strategy (arXiv:2509.24930).

When to Use

• Style specification from upstream pipeline is complete (or partially complete -- see Insufficient Data Handling)
• Need to produce a prompt that an LLM subagent can use to generate text in a target voice
• Translating psycholinguistic analysis (LIWC, Big Five, MDPI archetype, stylometrics, readability, CAT, rhetorical structure, speech acts) into generative instructions
• Building a voice-replication prompt for ANY text corpus, not limited to any specific platform
• Creating a testable, versioned persona prompt with verification criteria

When NOT to use: