Jonathon's AI Wiki

Marketing-Specific Prompt Patterns — Voice Preservation, Banned-Pattern Enforcement, Tone Calibration

9 min read

Source: wiki synthesis: Banned AI Patterns, Voice Profile Extraction, OmniPresence System (internal ai-video-content/ production articles, unpublished — see frontmatter sources: for paths; since 2026-07-09 those three are migration stubs, and their content lives in the work vault at weomarketly:playbooks/banned-ai-patterns, weomarketly:playbooks/voice-profile-extraction and weomarketly:projects/omnipresence-system, where the ban-list and five-category extraction claims were re-checked on 2026-09-29), and Module 1 — Prompts as Reusable Artifacts. 2026-09-29 field corroboration for Pattern 2: raw/30M_Writer_-_Never_write_AI_slop_again.md (Greg Isenberg with Nicolas Cole, youtube.com/watch?v=YuOSyRj3sXg).

The wiki has three well-developed, production-tested prompt patterns for making Claude write in someone else’s voice while suppressing generic “AI-sounding” output — but they’ve lived exclusively inside ai-video-content/ as dental-marketing production rules (OmniPresence’s script pipeline). Stripped of the dental framing, they generalize to any marketing, brand-voice, or ghost-writing prompt work. This article extracts the three patterns as general prompt-engineering technique.

Key Takeaways

  • Banned-pattern enforcement is constraint stacking made checkable. A flat list of forbidden words/phrases works better than “sound natural” because Claude can verify against a list but can’t verify against a vibe. The technique scales: 5 constraints in v1, 15+ by v3, each one earned from an observed failure.
  • Voice preservation works by extraction, not description. “Warm and plainspoken” is nearly useless as an instruction — it’s abstract, and Claude has to invent what it means. Pulling actual signature phrases, sentence-length patterns, and real stories from source material (a transcript, past emails, existing copy) and feeding those back as <examples> transfers voice with far higher fidelity than adjectives ever do.
  • Tone calibration is a named-reviewer self-critique step, not a rule. Asking Claude to read its own draft as a specific person with known pet peeves (“read this as Mel, who hates corporate jargon and rhetorical-question openers”) catches judgment-call violations that a checklist can’t enumerate — the things that are technically compliant but still read wrong.
  • These three patterns compose in a fixed order: extract the voice first (from real material), draft, then filter against the ban list, then critique in the reviewer’s voice. Running critique before extraction just produces confident-sounding genericness with nothing real to check it against.
  • Session isolation is the operational failure mode nobody’s constraint list catches. Voice bleed — one client’s or brand’s voice phrases leaking into another’s output — happens at the session level, not the prompt level. The fix is procedural (one identity loaded per session, fresh session on switch), not a prompt addition.

Pattern 1: The Checkable Ban List (Constraint Stacking)

The core move, generalized from Banned AI Patterns’s 99-phrase dental-marketing list: don’t tell Claude to “avoid sounding like AI” — give it an explicit, growable list of forbidden words and structural patterns it can check its own output against.

Two tiers of ban:

  • Word/phrase level — “diving into,” “unlock,” “leverage,” “game-changer,” “at the end of the day,” and similar generic-marketing-copy tells. These transfer to any domain; a WEO onboarding module uses an overlapping list: Module 1’s constraint-stacking example bans “streamline, leverage, world-class, game-changer, state-of-the-art” (two of the five phrases above). (Corrected 2026-09-29 from “the identical phrase set”.)
  • Structural-pattern level — this is the less obvious, higher-value half of the technique. “No X. No Y. Just Z.” sentence fragments, stacked short-punchy-fragment lists (“One visit. Done.”), standalone rhetorical-question openers (“The good news?”), and em-dash overuse are AI structural tells that survive a word-level ban list untouched. A generic word filter misses all of them; only naming the pattern catches it.

Why it outperforms “write naturally”: Claude can run a genuine check against an enumerated list (per Module 1’s validation-with-retry technique: “Output a <validation> block listing each banned phrase and whether you found it in your draft” — forcing the listing makes the check real instead of performative) but cannot meaningfully self-assess against an adjective like “natural.” The list compounds over time: start with 5 rules, add one every time a new failure mode is observed, and by the third production cycle you have 15+ rules earned the hard way rather than guessed upfront.

Pattern 2: Voice Extraction From Source Material

Voice Profile Extraction’s five-category framework — signature phrases, speech patterns, key stories, personality markers, recurring topics — is a general method for building a voice profile from any real transcript, email archive, or writing sample, not just a dental-practice interview:

  • Signature phrases: coined terms and metaphors the source repeats more than once. Repetition is the signal that a phrase is genuinely theirs, not a one-off.
  • Speech/sentence patterns: short-punchy vs. flowing, tag questions, contractions, where tone shifts from casual to serious.
  • Key stories: specific, real, told-with-energy examples — never fabricated, never paraphrased into genericness.
  • Personality markers: what they get animated about, what they’re proud of, what values surface unprompted.
  • Recurring topics: what they return to across the source material without being asked — this usually reveals what they actually want to be known for.

The extraction feeds directly into Module 1’s multi-shot-examples technique: 2–3 real on-voice examples in <examples> tags transfer sentence rhythm and implicit “what we don’t say” far better than any prose description of tone — “Examples teach what you do want; rules teach what you don’t” (Module 1’s wording), and the two belong in different sections of the same prompt. The generalization beyond dental marketing: any ghost-writing, brand-voice, or executive-communications prompt benefits from treating a real transcript or writing sample as the primary source, not a set of adjectives someone used to describe the person.

Field corroboration: “approved language” (added 2026-09-29). Nicolas Cole is a ghostwriter; Greg Isenberg’s intro says he “made over 30 million dollars writing on the internet” (raw/30M_Writer_-_Never_write_AI_slop_again.md). His account of ghostwriting is the same method, stated as a rule:

  • Ghostwriting is recombination. A ghostwriter doesn’t put “net new thinking in the client’s mouth”; they recombine the client’s “approved language,” meaning sentences the client has already written or said. Example: “a sentence from chapter one and a sentence from chapter two” of their book, stitched into something new.
  • AI should work the same way. “AI is taking approved language from things that you’ve written and creating net new combinations,” and “the larger your library grows… the more net new combinations you can create.”
  • Three content types. Cole sorts content into commodity, personality and original. What readers call “AI slop” is commodity content: sentences “not attributable to any one individual,” such as “the key to losing weight is eating healthy.”
  • Caveat. This is an operator’s opinion, and Cole sells a writing product built on it (Typeshare).

Applied to Pattern 2:^[inferred] keep a growing file of the client’s own sentences and stories, and feed from it. A one-off set of three examples is a snapshot; the library is what keeps the voice from drifting toward commodity phrasing.

Pattern 3: Self-Critique as a Named Reviewer Persona

The highest-leverage, least-obvious pattern. Module 1’s Technique 8 (self-correction) and the OmniPresence pipeline’s quality gate converge on the same move: after drafting, have Claude critique its own output as a specific named person with known pet peeves, not as a generic “check for quality” step.

Read your draft as if you were [Reviewer], [role]. [Reviewer]'s pet peeves:
[specific, concrete list — e.g. sentences that sound like a brochure,
generic opener patterns, hedging language, clinical jargon in casual answers].
List every issue you find, then produce the corrected version.

This catches what a rules-only validation pass misses: “this sentence is technically allowed but reads off-voice.” A banned-phrase list is binary (present/absent); a named-reviewer critique makes a judgment call and shows its reasoning, which a human can then agree or disagree with. The specificity of the pet-peeves list matters — “check for quality” produces a rubber-stamp pass; “Mel hates rhetorical-question openers and anything that sounds like a brochure” produces a real critique, because it gives Claude something concrete to test against rather than something to perform compliance with.

Generalizing Beyond Dental Marketing

None of the three patterns above are dental-specific — the domain material (gumline vs. gum line, Chicago Manual of Style numeral rules, “cheat code” as Dr. Browning’s coined phrase) is the input, not the technique. The technique is:

  1. Extract real voice material before writing anything (Pattern 2).
  2. Draft against that extracted voice.
  3. Filter the draft against an enumerated, growable ban list — words and structural patterns both (Pattern 1).
  4. Critique the filtered draft as a specific named reviewer with concrete pet peeves (Pattern 3).
  5. Keep the whole stack in one session per brand/client identity; never load two voice profiles in the same session.

Any team doing recurring brand-voice work — agency client scripts, executive ghost-writing, personal-brand content, customer-facing support macros — can lift this five-step stack directly. The dental-specific ban list and Chicago Manual of Style rules are one instantiation of step 3; a different brand needs its own list, built the same way (earned from observed failures, not guessed upfront).

  • weomarketly:playbooks/banned-ai-patterns — the source ban list and structural-pattern catalog this article generalizes (migrated to the internal work vault 2026-07-09).
  • weomarketly:playbooks/voice-profile-extraction — the source five-category extraction framework (migrated 2026-07-09).
  • weomarketly:projects/omnipresence-system — the two-layer (fixed structure + variable voice) production architecture these patterns were built inside (migrated 2026-07-09).
  • weomarketly:playbooks/mels-feedback-rules — the real-world reviewer persona Pattern 3 is modeled on (migrated 2026-07-09).
  • Module 1 — Prompts as Reusable Artifacts — the general constraint-stacking, multi-shot-example, and self-correction techniques these patterns are specific applications of.
  • Prompt Engineering Essentials — the foundational few-shot and self-correction techniques underlying Patterns 2 and 3.
  • Anti-AI Slop Guide — the design-focused (not copy-focused) counterpart ban list.
  • The Marketing Prompt Stack — Voice Patterns, Skill Packaging, and Evals

Try It

  1. Build a ban list for your own recurring writing task the way Banned AI Patterns did: start with 5 rules from memory, then add one every time you catch a new AI-sounding tell in review. Include at least one structural pattern (not just words) — “No X. No Y. Just Z.” fragments are the most common tell a word-only list misses.
  2. Pull 3 real writing samples from whoever’s voice you’re trying to match and extract signature phrases, sentence-length patterns, and one real story — feed those in as <examples>, not as adjectives describing their tone.
  3. Add a named-reviewer critique step to your next content prompt. Pick a real person (or persona) with known, specific pet peeves and have Claude critique its own draft as that person before producing the final version.
  4. Start an “approved language” file per voice. Collect sentences and stories the person has actually written or said (posts, book chapters, call transcripts), and add to it after every piece they approve. Tell the model to build from that file and to flag any sentence it can’t trace to it. The flagging rule is an extension of Cole’s method, not something he describes.

Graph View

Backlinks