subagent-instruction-operationalization

Inventory the completed analysis dimensions, select 3-5 typical few-shot examples from the corpus, then translate each dimension into prompt sections ordered by enforcement strength: role frame, complexity targets, syntactic fingerprint, emotional valence, and rhetorical structure. Finish with self-verification instructions and test against source material.

Few-shot examples are the strongest constraint, followed by measurable numerical targets like Flesch-Kincaid grade and sentence length ranges. Distributional constraints such as speech act ratios are weakly enforceable through instructions alone and rely primarily on examples.

Seven or more of the nine dimensions supports a full prompt. Four to six yields a partial prompt with documented gaps. One to three produces only targeted constraints, and zero dimensions means the upstream analyses must be run first.

Generate text on three different topics, then measure FK grade, sentence length, sentiment compound, and function-word rates against the source fingerprint. Passing means FK within plus or minus 1 grade, sentence length within 3 words, and sentiment within 0.15 of target.

Lexical and complexity constraints are strongly enforceable, but structural patterns drift and distributional targets like speech act ratios are only weakly enforceable without few-shot examples. One-shot style transfer accuracy ranges from roughly 67.6% to 94.7% depending on prompting strategy.

Do not use it when no style specification exists yet, when the goal is impersonation for deception or harassment, or when the target is a generic professional tone. Simple role prompts suffice for generic styles without specification-driven constraints.