Turning Voice Analysis Into System Prompt Instructions
Session 6.5 · ~5 min read
Documentation vs. Operating Commands
A voice fingerprint is documentation. A system prompt instruction is an operating command. The difference is the difference between a building blueprint and a construction order. "The building has large windows" (blueprint) becomes "Install 6-foot by 4-foot double-pane windows on the south face, spaced 8 feet apart" (construction order). The AI needs construction orders.
Most people include their voice analysis in prompts without translating it. They paste "I tend to use short sentences" into the system prompt. The AI reads this as a description, not a command. It may or may not act on it. Prescriptive instructions get followed. Descriptive observations get considered and sometimes ignored.
Every descriptive observation must become a prescriptive instruction. "I use fragments for emphasis" becomes "Use one sentence fragment per paragraph, positioned after a longer sentence, to create emphasis through contrast. Fragment length: 2-6 words." The more specific the instruction, the more consistently the AI follows it.
The Translation Table
Each voice analysis observation translates into one or more system prompt instructions. The table below shows the pattern.
| Voice Analysis (Descriptive) | System Prompt Instruction (Prescriptive) |
|---|---|
| "I use short sentences" | "Average sentence length: 12-16 words. Vary between 4-word fragments and 25-word compound sentences. Never write three long sentences in a row." |
| "I'm direct" | "Open every paragraph with a declarative statement. No questions. No 'it's worth noting.' State the point in the first sentence." |
| "I use humor" | "Include one moment of dry humor per section. Delivery: understated, not slapstick. Place it after a serious point, as a release valve." |
| "I don't hedge" | "Never use: arguably, perhaps, it could be said, it's important to note, some might say. State claims directly. If uncertain, say 'I don't know' rather than hedging." |
| "My metaphors come from manufacturing" | "When using metaphors, draw exclusively from manufacturing, production lines, and factory operations. Never use sports, war, or journey metaphors." |
| "I'm conversational" | "Use contractions. Address the reader as 'you.' Vary paragraph length between 1 and 4 sentences. Occasional one-sentence paragraphs for impact." |
The Translation Process
Work through your voice fingerprint section by section. For each observation, ask three questions.
'I use fragments for emphasis'"] --> B["Q1: What specific action
should the AI take?"] B --> C["A1: Use one fragment
per paragraph"] A --> D["Q2: How often
should it happen?"] D --> E["A2: Once per paragraph,
after a longer sentence"] A --> F["Q3: What are the
boundaries?"] F --> G["A3: Fragment length
2-6 words. Not at
paragraph start."] C --> H["Prescriptive instruction:
'Use one 2-6 word fragment
per paragraph, positioned
after a longer sentence.'"] E --> H G --> H style A fill:#222221,stroke:#c8a882,color:#ede9e3 style H fill:#222221,stroke:#6b8f71,color:#ede9e3
- What specific action should the AI take? Convert the observation into a verb. "I use fragments" becomes "Use fragments."
- How often or how much? Quantify. "Sometimes" is not useful. "Once per paragraph" or "in 30% of paragraphs" is.
- What are the boundaries? Define the limits. How long is a fragment? Where does it appear? What should it not be?
Testing the Instructions
After translating all observations, assemble the instructions into a system prompt and test. Generate 1,000 words using only these instructions (no few-shot examples). Then evaluate:
- Does the sentence length match your actual writing? Count words in 10 sentences.
- Are forbidden words absent? Search the output.
- Does the tone feel right? Read it aloud.
- Are the structural patterns present? Check paragraph openings, fragment placement, metaphor sources.
Where the output diverges from your voice, the instruction for that dimension is not specific enough. Refine it. Add examples. Add negative constraints ("never do X" is often more effective than "do Y"). Test again.
The Blind Test
The ultimate validation: have someone who knows your writing evaluate the output blind. Do not tell them it was AI-generated. Ask: "Does this sound like me?" If they say yes without hesitation, your instructions are working. If they hesitate or say "mostly," ask them to identify what feels off. That feedback maps directly to instructions that need refinement.
Further Reading
- Best Practices for Prompt Engineering, Anthropic
- Writing Style Analysis Using AI
- How to Humanize AI Text: Pro Writers Share Their Secrets, WriteHuman
Assignment
Take your voice fingerprint and translate every descriptive observation into a prescriptive system prompt instruction using the three-question method. Assemble the instructions into a system prompt. Generate 1,000 words using only these instructions (no examples). Have someone who knows your writing evaluate it blind. Can they tell it is AI? Where does it fail? Refine the instructions for those failures.