The Voice Clone Cliché: How to Audit LLM Custom Instructions to Strip Out Manufactured Executive Tone

The promise of using large language models (LLMs) for executive communication was supposed to be efficiency: feed a transcript or a few rough bullet points into an LLM, and get back a polished, professional post. Instead, the widespread reliance on standard system prompts has created a distinct, homogenized dialect.

This standardized dialect is characterized by predictable structural patterns, specific transition words, and an unearned enthusiasm. When multiple executives in the same industry rely on these default settings, their unique communication styles blend into a single, recognizable AI-generated voice.

To reclaim an authentic digital identity, founders and operators must audit their custom instructions in platforms like OpenAI’s ChatGPT and Anthropic’s Claude. This process requires identifying the default writing patterns of LLMs, cataloging individual voice metrics, and translating those metrics into concrete instructions.

The Default LLM Writing Signature

Large language models are trained to be helpful, polite, and structured. Left to their own defaults, they write with a specific signature that immediately signals machine generation to human readers.

This default signature relies on several identifiable elements:

• Symmetrical Structure: A typical three-to-five paragraph layout that begins with a broad setup, follows with bullet points containing bold lead-ins, and concludes with a neat summary.
• Predictable Transitions: Heavy reliance on transition words designed to guide readers logically but which rarely appear in spontaneous human speech.
• The Unearned Climax: A closing paragraph that attempts to elevate a tactical business observation into a profound lesson about the future of work or human connection.
• Superlative Adjectives: Frequent use of words that overstate the importance or complexity of a situation.

When auditing your current outputs, look for these structural tells. If your drafted posts consistently use structured lists with bold introductory phrases, or if they regularly conclude by framing a standard operational update as a journey, the LLM is overriding your natural communication style with its default safety and training constraints.

Step 1: Document Your True Verbal Fingerprint

Before writing new custom instructions, you must isolate how you actually communicate when you are not trying to write a formal post. The most effective way to capture this is through spoken speech rather than written drafts, as written drafts are often already filtered through what we believe we should sound like.

To establish a baseline, record yourself explaining a recent business decision or industry event to a colleague. Speak for three to five minutes without a script. Use a transcription tool to generate a raw text file of your speech.

Analyze the transcript for the following structural metrics:

Sentence Length Variance

Count the words in your sentences. Human speech is naturally uneven. A short, three-word sentence is often followed by a long, winded clause. LLMs prefer balanced, medium-length sentences. Note your natural variance. Do you speak in short, punchy fragments, or do you favor complex, multi-clause thoughts?

Contraction Density

Count how often you use contractions. In natural speech, humans use contractions almost exclusively. LLMs frequently default to formal, uncontracted words unless explicitly told otherwise.

Industry-Specific Jargon vs. Plain Language

Identify the specific shorthand you use to describe your work. Do you use technical terms naturally, or do you rely on analogies? Conversely, look for corporate buzzwords that you never actually say out loud and flag them for exclusion.

Qualification and Certainty

Note how you express doubt or certainty. Do you use phrases like “I think,” “historically,” or “in my experience,” or do you make direct, unprompted assertions?

Step 2: Write Negative Instructions to Break LLM Constraints

Most custom instructions fail because they only tell the AI what to do. LLMs are highly responsive to negative constraints—explicitly telling the model what not to do is often more effective than providing positive guidelines.

Create a dedicated “What to Avoid” section in your custom instructions. This section should target the specific linguistic habits that make AI text feel manufactured.

Ban Specific Transition Words and Clichés

Explicitly forbid the model from using transition words that do not belong in casual business communication. Common culprits include: * In conclusion * Furthermore * Moreover * Importantly * Indeed * Ultimately

Outlaw Bold Lead-ins in Lists

LLMs love formatting lists with bold introductory phrases. Instruct the model to write lists as natural, flowing sentences or simple bullet points without bold text or colon-separated headers.

Prohibit the “Wrap-up” Ending

Instruct the model to end the piece immediately after the final point is made. Explicitly state: “Do not write a concluding paragraph that summarizes the post or attempts to find a deeper, philosophical meaning in the topic.”

Step 3: Implement Positive Rules in ChatGPT and Claude

Once the negative constraints are established, translate your verbal fingerprint into positive instructions. ChatGPT’s “Custom Instructions” and Claude’s “Project Instructions” provide designated areas to store these rules permanently.

Structure your positive instructions using clear, non-negotiable directives:

# Writing Style Guidelines

- **Sentence Structure:** Use highly varied sentence lengths. Follow a 25-word sentence with a 3-word sentence. Use sentence fragments occasionally for emphasis.
- **Tone:** Conversational, direct, and slightly skeptical. Write as if explaining a concept to a peer over coffee.
- **Contractions:** Use contractions in every possible instance (e.g., use "don't" instead of "do not", "it's" instead of "it is").
- **Vocabulary:** Use plain, direct English. Avoid corporate jargon and decorative adjectives. If a word sounds like it belongs in a press release, replace it.
- **Punctuation:** Use em-dashes and parentheses to insert brief, conversational side-notes, just as they appear in natural speech.
- **Formatting:** Write in short paragraphs, rarely exceeding three sentences. Do not use bold lead-ins for bullet points.

Step 4: Run the Before-and-After Audit

To test the effectiveness of your new instructions, run a comparative test. Take a raw transcript or a set of rough bullet points and run them through two different chats: one with your new custom instructions active, and one in a default, clean session.

Compare the outputs side-by-side. Check for the survival of your natural phrasing. If the custom-instructed output still feels sterile or overly structured, continue to add negative constraints targeting the specific phrases or layouts that feel artificial.

By systematically stripping away the default patterns of the LLM, you can use generative tools to scale your output without sacrificing the individual voice that makes your perspective valuable to your audience.

This article was generated with the help of AI.