Building a Style Specification: The Document That Makes AI Match Our Voice

AI & Writing Part 3 of 4
Quick Takeaways
  • A style spec has six parts: voice, sentence form, word choice, layout, rhetorical moves, and samples
  • Claude handles tone and complex prose well. ChatGPT is strong with word lists and layout.
  • Specs fail when too vague ("be casual") or too strict ("use 15-word lines")

Stylometry gives us hard data. We get sentence length, word choice, and rhetorical moves. These are more precise than vague labels like "casual" or "bold."

But AI needs clear rules, not vague notes. A style spec bridges the gap. It turns our writing data into rules AI can follow.

This guide covers the full spec framework. We also walk through setup in Claude and ChatGPT.

The Six Parts of a Style Spec

A style spec is more than a list. It is a rule set. It covers tone, form, and word choice. Think of it as a style guide plus a manual.

Visual overview of the six components of a complete style specification: voice attributes, sentence patterns, vocabulary guidelines, structural preferences, rhetorical moves, and sample passages
The six components that make a style specification complete and actionable

1. Voice (Tone)

This part sets the core feel of our prose:

How Formal: Where we fall on the casual-to-formal scale

Example: "Warm but pro. Uses short forms. Skips jargon."

How We Show Skill: Cite facts, not our resume

Example: "Proof-based. Cite research, not credentials."

Tone with Readers: Peer-to-peer, not teacher-to-student

Example: "We talk with readers, not at them."

2. Sentence Form

This part logs our sentence habits:

Average Length: From our analysis (e.g., 15-20 words)

Mix: Simple, compound, and complex types

Example: "Mostly plain statements. Few fragments. Not many questions."

Marks: Dashes, colons, commas, and so on

3. Word Choice

This part maps our word habits:

Words We Use: Terms we pick most often

Example: "Say 'framework' not 'paradigm.' Say 'approach' not 'method.'"

Words We Skip: Buzzwords and clichés

Example: "Never: synergy, leverage, game-changer, deep dive"

How Concrete: Concrete terms or abstract ones?

4. Layout

This part sets how we shape our text:

Block Length: How long each block is (e.g., 3-5 lines)

Flow: How we order longer pieces

Example: "Problem, then proof, then fix, then use"

Bridges: Do we use clear linking words or let ideas flow on their own?

5. Rhetorical Moves

This part shows how we build our case:

How We Start: Lead with the problem, or define terms first?

Example: "Start with the real-world problem, not theory"

How We Use Proof: Weave it in or cite it formally?

Example: "Say 'Studies show X.' Not 'Smith (2020) found X.'"

What We Never Do: Moves we always skip

Example: "Never say what comes next. Just write it."

6. Sample Text (With Notes)

Add 2-3 excerpts with notes on what makes them "us":

[SAMPLE PASSAGE]
"The problem with most AI writing advice? It assumes writers want to sound
like everyone else. But distinctive voice isn't a luxury; it's the
difference between content people skim and content they remember."

[ANNOTATIONS]
- Opening question for engagement (use sparingly)
- Immediate answer (no suspense-building)
- Rhetorical contrast ("not X, but Y")
- Concrete outcome focus (remember vs. generic "engage")
- Sentence variety: 7 words, then 17 words

Setup: Claude Custom Styles

Claude has a clean way to save style specs.

Screenshot showing Claude's custom styles interface with the five-step implementation process: prepare, access, structure, test, and iterate
Claude's custom styles provide persistent, named style specifications

Step 1: Prep the Spec

Before we paste it in, make sure it is:

  • Full: All six parts done
  • Clear: Real samples, not vague notes
  • Short: 2-3 pages max (AI has input caps)

Step 2: Access Custom Styles

  1. Open Claude.ai or Claude desktop app
  2. Navigate to Settings → Custom Styles
  3. Click "Create New Style"

Step 3: Add the Spec

Pick a clear name (e.g., "Pro_Analysis_Voice"). Format it for AI:

VOICE ATTRIBUTES:
- Formality: Professional but approachable; uses contractions
- Authority: Evidence-based, not credential-based
- Reader relationship: Peer collaboration

SENTENCE PATTERNS:
- Average length: 18 words
- Variety: Mix declarative with occasional fragments for emphasis
- Punctuation: Semicolons for related ideas; em-dashes sparingly

VOCABULARY:
- Prefer: framework, approach, patterns, specific
- Avoid: synergy, leverage, paradigm, game-changer
- Level: Educated general audience, minimal jargon

STRUCTURE:
- Paragraphs: 3-5 sentences
- Organization: Problem → Evidence → Solution
- Transitions: Implicit flow over explicit connectors

RHETORICAL MOVES:
- Introductions: Lead with reader's problem
- Evidence: Integrated, not cited formally
- Never: Announce structure, use rhetorical questions excessively

SAMPLE (annotated):
[Include strongest example with annotations]
Preview of a completed style specification template showing all six sections filled in with specific, measurable guidance
A complete style specification provides concrete, measurable guidance across all six dimensions

Step 4: Test It

Write 3-5 test pieces. Check each one:

  • Do sentence lengths match ours?
  • Are the right words used?
  • Does the layout feel right?
  • Does it sound like us?

Step 5: Refine

Fix what went wrong:

  • Too formal? Add casual samples.
  • Wrong words? Grow the word lists.
  • Bad layout? Add more layout rules.

Log each miss. When AI keeps failing at one thing, add a rule for it.

Setup: ChatGPT Custom Rules

ChatGPT uses two text fields, not named styles. The idea is the same. We just split the spec in two.

"What should ChatGPT know?"

Put in:

  • What we write and for whom
  • Our voice and tone
  • Key word picks

"How should ChatGPT respond?"

Put in:

  • Sentence length rules
  • Layout and flow rules
  • Rhetorical moves and bans
  • One sample with notes

Note: ChatGPT has a stricter text cap than Claude. We may need to keep only the most unique parts of our voice.

Claude vs. ChatGPT

Tests on both platforms show different strengths:

Claude Strengths ChatGPT Strengths
Keeps complex sentence forms Sticks to word lists well
Holds tone well Good at layout rules
Reads long specs without cutting off Good at fixed patterns
Styles save across chats —

Both struggle with:

  • Subtle voice traits that are hard to spell out
  • Style shifts based on context
  • Staying on track in long outputs
Which One to Pick

If our voice leans on rhythm and tone, Claude tends to do better. If our voice is more about set words and fixed layout, ChatGPT may work well.

Traps to Dodge

Too Vague vs. Too Strict

Too Vague:

"Be casual and fun"

Flaw: AI reads it a new way each time

Fix: "Use short forms 60-70% of the time"

Too Strict:

"Always use 15-word lines"

Flaw: Output sounds robotic

Fix: "Aim for 15-20 words. Range: 8-30."

Too Much vs. Too Little

Too Much:

A rule for every word choice

Flaw: AI can't rank what matters most

Fix: Cover only what is unique to us

Too Little:

Tone only, no form rules

Flaw: AI has no clear guide

Fix: Add both voice and form rules

Stale Specs

Our voice shifts. Our spec should too.

  • Check it every few months
  • Update after big style shifts
  • Test against recent work, not old drafts

One Spec for All

Using one spec for all contexts is a trap:

  • Flaw: Social posts sound like essays
  • Fix: Make a few versions (Pro, Social, School)
  • Keep our core voice. Adjust tone and layout.
Action Steps
  1. Grab our data: Pull the patterns from Article 2
  2. Fill all six parts: Voice, sentences, words, layout, rhetorical moves, and samples
  3. Pick a tool: Claude for tone. ChatGPT for word rules.
  4. Test 3-5 pieces: Write on different topics to check
  5. Fix what fails: Note each miss. Refine the spec.

What Comes Next

Article 2 gave us the data. This article gave us the spec format. Now we need a full workflow from idea to final draft.

Article 4 is about testing AI output. We look for where AI fails. Then we fix the spec. This feedback loop is the last piece.