Voice Profile Trainer
What This Skill Does
Generic AI content has a tell. It’s competent, readable, and structurally sound — and it sounds like everyone else. The “stochastic average problem” is real: without a deliberate voice layer, AI writing converges toward the statistical center of its training data. That center is not your voice.
This skill reverses that. It analyzes your authentic writing samples, identifies the specific psycholinguistic patterns that make your voice yours, and encodes them into a reusable prompt block — a Voice Profile — that can be prepended to any content generation request. Once built, the profile dramatically reduces the gap between AI output and your authentic voice on the first pass, cutting revision time and producing content that actually represents you.
The deeper principle: Voice profile is not about style constraints. It is about IP recovery. The way you write — your sentence rhythms, your argumentative moves, your relationship to concrete examples — encodes your cognitive fingerprint. That fingerprint is intellectual property. This skill recovers it from your existing writing and makes it portable.
UP=IP Principle (from Feb 19 session): Your editing pattern — what you consistently move toward and away from when revising AI output — encodes your intellectual identity. The edit-away examples are more diagnostic than the approved examples. What you reject tells you more about your voice than what you accept.
Score: 89/100 across 6 sessions. Second-highest scoring pattern in the vault.
Quick Start
If you want to build your profile fast:
- Gather 10-12 writing samples you consider “genuinely me” (see Stage 1 below for guidance on what qualifies).
- Gather 5-10 AI outputs you received and significantly corrected — the “edit-away” examples.
- Say: “Build my voice profile from these samples.”
- Review the extracted pattern report and confirm before the profile is finalized.
- Receive your named, versioned Voice Profile prompt block (200-400 words), ready to use immediately.
Building Your Profile
Stage 1 — Sample Collection Intake
What: Collect two distinct sets of material.
Set A — “Genuinely Me” samples (10-12 pieces):
Writing that represents your authentic expression. The test for inclusion: you wrote it without heavily editing AI output, and when you read it back, it sounds like you. Examples:
- Published articles or blog posts you wrote from scratch
- Emails to clients or colleagues where you were thinking clearly and writing freely
- LinkedIn posts you’re proud of
- Coaching notes or session summaries you wrote in the moment
- Newsletters, letters, or proposals from your best work periods
Aim for variety across formats, but all from the same practitioner. The goal is to find the invariant patterns — what stays consistent across email, article, and social post. Include pieces from different time periods if possible; durable patterns are more reliable than recent quirks.
Set B — “Edit-Away” samples (5-10 pieces):
AI outputs you received and meaningfully corrected. Not samples you discarded entirely — samples you worked with and edited. The edits you made are the data. For each, ideally provide:
- The original AI output
- Your edited version
- (Optional) A brief note on what felt wrong before you corrected it
If you don’t have saved examples, describe 3-5 types of AI output that consistently bother you and what you change about them. The pattern still emerges.
Why Set B is more diagnostic than Set A: Your approved writing contains everything you do — including the things you do by accident or habit. Your edited AI output shows what you do intentionally — the conscious corrections reveal your standards. The edit-away examples are a direct readout of your editorial identity.
Stage 2 — Pattern Extraction
What: Analyze both sample sets for the following dimensions.
Sentence length distribution and rhythm:
- What is the distribution of short (under 10 words), medium (10-20 words), and long (20+ words) sentences?
- Does the practitioner use short sentences for emphasis at the end of a paragraph? At the start?
- Target Flesch-Kincaid readability range: 6-8 is practical for entrepreneur/executive audiences. Scores below 6 suggest overly complex prose; scores above 10 suggest too casual or fragmented.
- Note: FK score is a proxy, not a target. The goal is to match the practitioner’s actual rhythm, not to optimize a number.
Structural preferences:
- Does the practitioner lead with story (a specific scene or moment), argument (the claim stated plainly up front), or framework (a named mental model introduced first)?
- Do they use numbered lists? If so, at what density, and in what contexts?
- How do they open and close? What are the first and last moves?
Vocabulary register:
- Casual or formal? Where on the spectrum does the practitioner naturally sit?
- Jargon density: Do they use industry terms freely, avoid them, or define them when used?
- Metaphor types: Do they favor sports analogies, military analogies, nature analogies, mechanical analogies, none?
- Characteristic word choices: Words they reach for repeatedly that feel distinctly theirs
AI telltales to flag and remove: These are patterns that reliably identify AI-generated content and that most practitioners would edit away. Flag any that appear in the approved samples (they may need to be unlearned) and treat them as hard filters for the Voice Profile:
- Excessive list structures (everything broken into threes and sixes)
- Transition words: “delve,” “nuanced,” “tapestry,” “certainly,” “it’s worth noting,” “in today’s [X] landscape”
- Hedging phrases: “it’s important to remember,” “one might argue,” “this is a complex topic”
- Passive voice used where active voice was available
- Over-explanation: restating the conclusion at the end of a paragraph that already made it
- Preamble: starting an answer with a restatement of the question
- False balance: “on one hand… on the other hand” when the practitioner actually has a view
The practitioner’s signature moves:
- Recurring argumentative structures (e.g., “here’s what most people do wrong, here’s why, here’s what to do instead”)
- Favorite types of evidence (personal anecdote, client story, named thinker, statistics, first principles)
- Ways they handle objections — directly, preemptively, or not at all?
- The emotional register they write in — warm, analytical, provocative, pragmatic?
Stage 3 — Cognitive Fingerprint
What: From the edit-away examples, identify the consistent pattern of what the practitioner moves toward when they correct AI output.
Why this is the most important stage: The edit-away corrections are intentional choices. They reveal standards that the practitioner holds but may not have articulated. Common directions practitioners move in:
- Toward more directness (removing hedges, cutting preamble, stating the view plainly)
- Toward more story (replacing abstract claims with specific scenes or named examples)
- Toward more concrete specificity (replacing “many leaders” with “the executive at the session last Tuesday”)
- Toward shorter paragraphs or sentences (breaking up dense prose)
- Toward a stronger closing (replacing vague endings with a clear final move)
- Away from lists (converting bullet points back to prose)
- Away from formal register (replacing “it is essential to consider” with “you have to think about”)
The cognitive fingerprint is the 2-4 most consistent patterns across all the edit-away examples. These are the practitioner’s non-negotiables — the things their voice insists on regardless of topic or format.
Name the fingerprint explicitly. For example: “Lou’s fingerprint: direct openers, concrete specificity, short punchy closers, no lists in arguments.” The name makes it portable and memorable.
Stage 4 — Style Guide Generation
What: Produce a named, versioned style guide document that encodes all findings.
Format: [Name]'s Voice Profile v[X.X]
The style guide includes:
- The cognitive fingerprint (2-4 core patterns)
- Sentence rhythm targets (FK range, short/medium/long distribution)
- Structural preferences (how to open, how to close, list usage rules)
- Vocabulary register notes (jargon policy, metaphor preferences)
- AI telltales blacklist (specific words and phrases to avoid)
- Characteristic vocabulary (words and phrases that feel authentically theirs)
- Signature argumentative moves (2-3 recurring structures to emulate)
This document lives in the vault as a persistent asset. It should be stored at:
wiki/mastermind/skills/voice-profile-trainer/[name]-voice-profile-v[X.X].md
Stage 5 — Authentication Threshold
What: Establish a voice similarity check — a short set of questions to self-assess whether a draft has been adequately filtered through the voice profile.
Why it matters: Without an explicit threshold, the style layer becomes aspirational rather than functional. A draft that “mostly” sounds like the practitioner but still has 5 AI telltales in it is not ready. The authentication check makes the standard concrete.
Practical threshold (from Kasimir’s proposal, Oct 2025): 70-80% similarity before output is accepted for delivery. This is not a mathematical score — it’s a checklist-based assessment. Below is the standard template.
Voice Authentication Checklist:
- Cognitive fingerprint patterns are present (check each named pattern from Stage 3)
- No AI telltales from the blacklist appear in the draft
- Sentence rhythm matches the practitioner’s target distribution (spot-check 3 paragraphs)
- Opening move matches the practitioner’s structural preference
- Closing move matches the practitioner’s signature close
- Vocabulary register is consistent (no sudden formality shifts or jargon anomalies)
- Lists are used per the practitioner’s rules (or not used, if they avoid them)
Passing 5-6 of 7 checks = 70-80% threshold met. If fewer than 5 pass, return the draft to the style layer for another pass before delivery.
Stage 6 — Output: The Voice Profile Prompt Block
What: A 200-400 word instruction set, formatted as a system prompt block, that when prepended to any writing prompt brings AI output significantly closer to the practitioner’s authentic voice on the first pass.
Format:
## [Name]'s Voice Profile v[X.X]
When writing for [Name], apply these voice parameters:
**Cognitive Fingerprint (non-negotiable):**
[List the 2-4 core patterns identified in Stage 3]
**Sentence Rhythm:**
[Target FK range and short/medium/long distribution guidance]
**How to Open:**
[Structural preference: story-first, argument-first, or framework-first — with a brief example of the opening move]
**How to Close:**
[The practitioner's characteristic closing move]
**Vocabulary Register:**
[Casual/formal position, jargon policy, metaphor types to use or avoid]
**Hard Removes (AI Telltales):**
[Blacklist — specific words and phrases never to use]
**Characteristic Vocabulary:**
[Words and phrases that feel authentically theirs — use naturally, not excessively]
**Signature Move:**
[One recurring argumentative structure to emulate when it fits]
Using the prompt block: Paste it at the top of any content generation request. It works as a system-level filter — the writing engine reads the parameters and applies them throughout the draft. It does not need to be re-explained; it simply prepends.
Example usage with the transcript-to-content-pipeline skill:
[Voice Profile block] Now produce a LinkedIn post from the following nugget: [nugget]
Sample Voice Profile Template
Below is a filled-in example based on the patterns identified in the PowerUp Coaching mastermind sessions. Use as a reference — not as a default.
## Lou's Voice Profile v1.0
When writing for Lou, apply these voice parameters:
**Cognitive Fingerprint (non-negotiable):**
- Direct openers: state the claim in the first sentence, no preamble
- Concrete specificity: replace abstractions with named examples, specific numbers, or real scenes
- Short punchy closes: the last sentence of any piece should land, not trail off
- No lists in arguments: when making a case, use prose; lists are for reference only
**Sentence Rhythm:**
- Target Flesch-Kincaid: 6-8
- Use short sentences (under 10 words) for emphasis — one per paragraph maximum
- Medium sentences carry the argument; long sentences are used sparingly for complexity
**How to Open:**
- Argument-first: lead with the claim. "Most coaches are optimizing for the wrong thing."
- Not: "In today's rapidly evolving coaching landscape, many practitioners are beginning to realize..."
**How to Close:**
- A single declarative sentence that names what the reader should do or think next
- Not a summary. Not a call to action with a button. A final move.
**Vocabulary Register:**
- Casual-professional: smart but not formal, precise but not academic
- Low jargon density unless the audience is assumed to know the term
- Favored metaphors: engineering and systems thinking ("leverage point", "feedback loop")
- Avoid: sports analogies, military metaphors
**Hard Removes:**
delve, nuanced, tapestry, certainly, it's worth noting, in today's X landscape,
it's important to remember, one might argue, on one hand... on the other hand,
[any sentence starting with "This is"]
**Characteristic Vocabulary:**
leverage, compound, judgment, practitioner, knowledge entrepreneur, canonical, signal
**Signature Move:**
"Here's what most [audience] do — here's why it doesn't work — here's the shift."
Used in 2-3 pieces per month, not every piece.
Using Your Profile
Once built, the Voice Profile prompt block is a portable asset. Use it:
With the transcript-to-content-pipeline skill: Mention that a voice profile exists at the start of any content run. The pipeline will apply it automatically in Stage 6.
With standalone writing requests: Paste the voice profile block at the top of any writing prompt:
[Voice Profile block] Write a LinkedIn post about [topic], using the following key idea: [idea]
With Claude or any AI writing tool: The profile block works as a system prompt segment. Paste it into the system prompt field if available, or at the top of the user prompt if not.
With other practitioners: The profile is named and versioned, which means it can be shared selectively. A practitioner sharing their voice profile with a writing collaborator is sharing their IP — treat it accordingly.
Updating Your Profile
A voice profile is not static. Voices evolve, and practitioners’ editorial standards sharpen over time.
When to update:
- After a significant body of new writing that feels more authentically you than the original samples
- When you notice consistent patterns in your edits that aren’t captured in the current profile
- After working with the profile for 3-6 months and finding specific elements that don’t ring true
- When your audience or publishing context shifts significantly (e.g., moving from LinkedIn-primary to newsletter-primary)
How to update:
- Collect 5-8 new “genuinely me” samples and 3-5 new edit-away examples
- Run them through Stages 2-3 (Pattern Extraction and Cognitive Fingerprint) only
- Compare findings against the existing profile — identify what has shifted
- Update the specific elements that have changed; preserve elements that remain stable
- Increment the version number: v1.0 → v1.1 for minor updates, v1.0 → v2.0 for significant shifts
- Keep the previous version in the vault — voice evolution is itself an interesting record
What not to change: Cognitive fingerprint elements that have been stable across multiple versions are likely deep patterns. Be slow to modify them. Changing a fingerprint element because one draft felt off is premature — change it when a pattern of multiple drafts and multiple editing passes consistently points in a new direction.
Common Pitfalls
Collecting samples from your best-edited work only: If all your samples are polished final drafts, the profile may over-index on your formal register and miss the warmth and directness of your unguarded writing. Include emails, coaching notes, and working documents as well as published pieces.
Skipping the edit-away examples: The “genuinely me” samples alone produce a descriptive profile of what you do. The edit-away examples produce a prescriptive profile of what you insist on. The second is more powerful and more actionable. If you have no saved AI outputs to use, dedicate 30 minutes to prompting AI to write about your topic area and then editing the output — the edits will reveal more than an hour of introspection.
Treating the profile as a style straitjacket: The profile should increase the probability that first-pass AI output sounds like you — it should not prevent the AI from writing anything that doesn’t fit the template exactly. Apply it as a filter and lens, not as a set of mandatory rules. The practitioner’s editorial judgment in review always overrides the profile.
Letting the profile go stale: A profile built entirely from 2023 writing applied to 2026 content will produce a slightly off-brand result. Schedule a profile review every 6 months. It takes less than an hour and the delta is worth capturing.
Not naming the cognitive fingerprint: An unnamed fingerprint is easy to forget and hard to defend. The name makes it portable: “This doesn’t pass Lou’s fingerprint — no concrete specificity, closes with a summary instead of a landing point.” Named patterns can be invoked in conversation; abstract patterns cannot.
Connected Skills
transcript-to-content-pipeline— The primary consumer of the voice profile. The pipeline applies the profile in Stage 6 (Style Layer).anthropic-skills:aimm-writing-team— The writing team can accept a voice profile block as a system-level constraint for any content run.aimm:cognitive-fingerprint— A related skill for exploring the practitioner’s deeper intellectual identity patterns beyond writing style.
Version History
- v1.0 — Initial pattern identified, Sep 2025 sessions
- v1.5 — Edit-away examples formalized as primary diagnostic input, Nov 2025
- v2.0 — Authentication threshold (70-80% similarity check) added; UP=IP principle named, Feb 2026
- v2.5 — Voice Profile prompt block format standardized; versioning protocol added, Mar 2026
- v3.0 — Current version. Cognitive fingerprint naming protocol added; update cadence guidance formalized, Apr 2026
Score: 89/100 across 6 sessions. Contributions: Lou, Kasimir, Bally, PowerUp Coaching mastermind group.