05 Voice Profiles

Voice Profiles

Voice profiles define how Ghost Writer sounds. Each profile specifies tone, vocabulary, structural habits, and perplexity/burstiness targets. You can use built-in profiles or create custom ones from your own writing samples.

---

The 4 Built-in Profiles

john-williams

Attribute	Value
Background	15+ year paid media veteran, Google Ads specialist, football coach mentality
Tone anchors	Direct and opinionated; speaks from specific experience; uses coaching/football analogies; comfortable with edge; mixes technical depth with accessibility
Domain terms	ROAS, PMax, SQOS, brand vs non-brand, impression share, wasted spend
Signature phrases	"here's the thing", "I've seen this blow up", "the data doesn't lie", "let me be real"
Avoid	synergy, leverage, holistic, paradigm, delve
Starts with conjunctions	Yes
Uses fragments	Yes
Parenthetical asides	Frequent
Avg sentence length	16 words
Sentence length stdev	9
Perplexity target	30–45
Burstiness	High

---

agency

Attribute	Value
Background	Digital marketing agency, data-driven, case study heavy
Tone anchors	Professional but not stiff; data-backed claims; case study references; ROI-focused
Domain terms	KPI, conversion rate, attribution, funnel, pipeline
Signature phrases	"the data shows", "in our experience", "across our client base"
Avoid	synergy, move the needle, circle back
Starts with conjunctions	No
Uses fragments	No
Parenthetical asides	Occasional
Avg sentence length	18 words
Sentence length stdev	7
Perplexity target	35–50
Burstiness	Medium

---

technical

Attribute	Value
Background	Engineer, developer, precise specifications
Tone anchors	Precise; specification-heavy; code-aware; no fluff
Domain terms	API, endpoint, payload, schema, latency
Signature phrases	"the implementation is", "under the hood", "the architecture"
Avoid	game-changer, revolutionary, amazing
Starts with conjunctions	No
Uses fragments	Yes
Parenthetical asides	Rare
Avg sentence length	14 words
Sentence length stdev	6
Perplexity target	40–60
Burstiness	Medium

---

casual

Attribute	Value
Background	Friendly peer, approachable, social-media-native
Tone anchors	Conversational; emoji-light; fragment-heavy; speaks like a friend
Domain terms	(none)
Signature phrases	"honestly", "look", "real talk", "not gonna lie"
Avoid	furthermore, consequently, utilize
Starts with conjunctions	Yes
Uses fragments	Yes
Parenthetical asides	Frequent
Avg sentence length	12 words
Sentence length stdev	11
Perplexity target	25–40
Burstiness	High

---

Creating Custom Profiles

Voice Import Process

1. Paste samples — Provide 2–5 writing samples (min 50 chars each, recommended 3+)
2. Statistical analysis — Engine computes:
   • Sentence length distribution (mean, median, stdev, min, max)
   • Vocabulary (TTR, hapax ratio, domain terms, signature phrases)
   • Structural habits (conjunction starters, fragments, parenthetical frequency)
   • Burstiness target (high/medium/low from stdev)
   • Perplexity proxy range
3. AI enrichment — GPT analyzes samples for:
   • Tone anchors (4–6 descriptive phrases)
   • Signature phrases (3–5 recurring expressions)
   • Writing style summary
4. Fingerprint extraction — Combined stats + AI output
5. Save — Profile stored (requires KV/CONTEXT binding)

What Gets Extracted

Category	Extracted Data
Sentence length	Mean, median, stdev, min, max
Vocabulary	TTR, hapax ratio, domain terms (top 20 by frequency), signature phrases (2-grams ≥2 uses)
Structural habits	Conjunction rate, fragment rate, parenthetical frequency (rare/occasional/frequent)
Tone anchors	4–6 phrases from AI analysis
Avoid list	Words from standard avoid list not used in samples
Perplexity target	Proxy range from sentence stats
Burstiness target	high (stdev ≥9), medium (≥5), low otherwise

API: Analyze Only

POST /api/writing-agent-voice
{
  "action": "analyze",
  "samples": [
    "Your first writing sample...",
    "Your second sample...",
    "Your third sample..."
  ]
}

API: Analyze and Save

POST /api/writing-agent-voice
{
  "action": "save",
  "name": "my-voice",
  "samples": [
    "Sample 1...",
    "Sample 2...",
    "Sample 3..."
  ]
}

Requires CONTEXT KV binding for storage.

---

Best Practices

• Use 3+ samples — More samples improve fingerprint accuracy
• Diverse content — Mix post types (LinkedIn, email, blog) for general voice
• Consistent author — All samples from same writer
• Minimum length — Each sample ≥50 chars; 500+ chars recommended