Style Specification Building

Building the Style Specification

Overview

Translate all prior analysis outputs into a single, implementable style specification by mapping each finding to a concrete writing constraint, resolving conflicts between analyses, and producing both a numeric profile and a prose description usable as an LLM system prompt or human writing guide. The core principle: every constraint in the specification must trace to a specific analysis finding with a stated confidence level, and no analysis finding should appear in the specification without being translated into an actionable constraint. Raw analysis outputs are observations; the style specification is a set of instructions. This skill performs the translation.

Why this is the penultimate step: All upstream analyses (personality profiling, stylometric fingerprinting, psycholinguistic categorization, readability metrics, rhetorical structure, accommodation patterns, archetype classification, register variation, speech act distribution, audience patterns, and persona framing) produce findings in their own frameworks. Without a specification-building step, those findings remain siloed observations. The style specification resolves them into a single coherent document that the final step (voice prompt engineering or human style guide creation) can consume directly.

Design influences: Nielsen Norman Group's four dimensions of tone (formal/casual, serious/funny, respectful/irreverent, enthusiastic/matter-of-fact); the COSTAR prompt framework (Context, Objective, Style, Tone, Audience, Response); Mailchimp and Google voice-and-tone documentation patterns; meta-analytic evidence that LLMs follow 5-12 well-prioritized constraints more reliably than 30+ undifferentiated rules.

When to Use

• Multiple linguistic analyses have been completed and need consolidation into one voice profile
• Building an LLM system prompt that replicates a specific writing voice
• Creating a human-readable style guide from quantitative analysis outputs
• Need to resolve contradictions between different analyses (e.g., personality suggests formality but stylometrics show casual function-word patterns)
• Preparing the penultimate deliverable before final prompt engineering or style guide assembly
• Converting a set of analysis reports into actionable writing constraints

When NOT to use:

• Fewer than 3 upstream analyses are complete (see Insufficient Data Handling)
• No analysis has been performed yet (run analyses first; this skill consumes their outputs)
• Goal is to analyze text (this skill synthesizes existing analyses, not raw text)
• The specification is for a team or brand voice with multiple contributors (this produces a single-author voice specification)
• Analysis outputs are from different authors being merged (this maps one voice, not a composite)

digraph spec_decision {
    rankdir=TB;
    "3+ analyses\ncomplete?" [shape=diamond];
    "Single author\nvoice?" [shape=diamond];
    "Analyses cover\n3+ dimensions?" [shape=diamond];
    "Proceed with\nfull specification" [shape=box];
    "Minimal specification\nwith heavy caveats" [shape=box];
    "STOP: Run more\nanalyses first" [shape=box];
    "STOP: Not designed\nfor multi-author" [shape=box];

    "3+ analyses\ncomplete?" -> "Single author\nvoice?" [label="yes"];
    "3+ analyses\ncomplete?" -> "STOP: Run more\nanalyses first" [label="no"];
    "Single author\nvoice?" -> "Analyses cover\n3+ dimensions?" [label="yes"];
    "Single author\nvoice?" -> "STOP: Not designed\nfor multi-author" [label="no"];
    "Analyses cover\n3+ dimensions?" -> "Proceed with\nfull specification" [label="yes"];
    "Analyses cover\n3+ dimensions?" -> "Minimal specification\nwith heavy caveats" [label="no, redundant\ncoverage"];
}

Quick Reference: The Analysis-to-Constraint Mapping

Each upstream analysis maps to specific constraint categories. This table is the backbone of the specification.

Workflow

Copy this checklist and track progress:

Style Specification Building Progress:
- [ ] Step 1: Inventory available analyses and assess coverage
- [ ] Step 2: Extract key findings from each analysis report
- [ ] Step 3: Translate each finding into a constraint (with confidence)
- [ ] Step 4: Detect and resolve cross-analysis conflicts
- [ ] Step 5: Prioritize constraints by confidence and impact
- [ ] Step 6: Build the numeric profile
- [ ] Step 7: Write the prose specification
- [ ] Step 8: Validate specification implementability
- [ ] Step 9: Write findings to docs/analysis/24-style-specification.md

Step 1: Inventory Available Analyses and Assess Coverage

Before building the specification, determine which analyses are available and what dimensions they cover. Not all analyses need to be complete -- the specification handles missing inputs gracefully.

Coverage assessment matrix:

Minimum viable specification: At least 3 analyses covering at least 3 different dimensions. A specification built from fewer than 3 analyses or covering fewer than 3 dimensions should be labeled "partial specification" and must document which constraint categories have no evidential basis.

Dimension classification:

Step 2: Extract Key Findings from Each Analysis Report

For each available analysis, extract the findings that will translate into constraints. Read each report (from docs/analysis/) and record:

1. The finding (what was measured and what value was observed)
2. The confidence level (as stated in the original report: high, moderate, low, or indeterminate)
3. The evidence basis (corpus size, number of samples, statistical measure)
4. Any caveats the original analysis flagged

Extraction format per analysis:

### [Analysis Name] (Report: docs/analysis/NN-name.md)

| Finding | Value/Range | Confidence | Evidence Basis | Caveats |
|---------|-------------|------------|----------------|---------|
| [metric] | [value] | [H/M/L] | [basis] | [caveats] |

Critical rule: Do NOT reinterpret or re-analyze raw data at this step. Accept the upstream report's findings and confidence levels as given. If an upstream report flagged low confidence, preserve that flag. The specification is a translation layer, not a re-analysis.

Step 3: Translate Each Finding into a Constraint

This is the core translation step. For each extracted finding, produce a concrete writing constraint using the mapping table from Quick Reference.

Translation rules:

1. Use ranges, not exact values. A finding of "FK grade 12.3" becomes a constraint of "FK grade 11-13" (typically +/- 1 grade level). A finding of "article rate 6.7%" becomes "article rate 6.2-7.2%" (typically +/- 0.5 percentage points for function-word rates).

2. Scale constraint precision to confidence:

3. Use constraint verbs that match confidence:

4. Every constraint must cite its source analysis. Format: [constraint text] (Source: [analysis name], confidence: [H/M/L])

Translation examples per analysis type:

From MDPI Archetype (e.g., HHL classification):

• Finding: "Classified as HHL (High Score, High Sentiment, Low Toxicity) with typicality 0.87"
• Constraint: "Emotional register MUST maintain positive-to-neutral valence. Toxicity markers (profanity, aggressive rhetoric) MUST be avoided. Arousal level SHOULD be moderate -- enthusiastic but measured. (Source: MDPI Archetype, confidence: High)"

From Big Five Openness (e.g., score 72 +/- 12):

• Finding: "Openness score 72 (High), confidence band +/- 12"
• Constraint: "Vocabulary SHOULD include sophisticated/abstract terms. Long words (6+ chars) SHOULD constitute 18-24% of text. Type-token ratio SHOULD target 0.45-0.55 on 1000-word windows. Intellectual framing and tentative language ('perhaps', 'consider') are appropriate. (Source: Big Five Openness, confidence: Moderate)"

From Readability/Lexical Diversity (e.g., FK 12.3, MTLD 94):

• Finding: "FK grade 12.3, Coleman-Liau 13.1, MTLD 94, MATTR-50 0.72"
• Constraint: "Readability MUST target FK grade 11-14 (consensus range). MTLD MUST exceed 80. Average sentence length SHOULD be 18-22 words. Average word length SHOULD be 4.8-5.4 characters. (Source: Readability/LD, confidence: High)"

Step 4: Detect and Resolve Cross-Analysis Conflicts

Different analyses can produce contradictory constraints. This step identifies and resolves them.

Common conflict patterns:

Conflict resolution protocol:

digraph conflict_resolution {
    rankdir=TB;
    "Two analyses\nconflict?" [shape=diamond];
    "Do they measure\nsame feature?" [shape=diamond];
    "Which has\nhigher confidence?" [shape=diamond];
    "Not a true conflict:\ndifferent features" [shape=box];
    "Higher-confidence\nanalysis sets hard\nconstraint; lower\nbecomes soft preference" [shape=box];
    "Equal confidence:\nuse range that\nencompasses both" [shape=box];
    "Document both and\nstate cannot resolve" [shape=box];

    "Two analyses\nconflict?" -> "Do they measure\nsame feature?" [label="yes"];
    "Do they measure\nsame feature?" -> "Not a true conflict:\ndifferent features" [label="no, different\nfeatures"];
    "Do they measure\nsame feature?" -> "Which has\nhigher confidence?" [label="yes, same\nfeature"];
    "Which has\nhigher confidence?" -> "Higher-confidence\nanalysis sets hard\nconstraint; lower\nbecomes soft preference" [label="one is\nclearly higher"];
    "Which has\nhigher confidence?" -> "Equal confidence:\nuse range that\nencompasses both" [label="similar\nconfidence"];
}

Documenting resolutions: Every conflict MUST be documented in the specification with:

1. The conflicting findings (with sources)
2. Why they conflict
3. How the conflict was resolved
4. What evidence supported the resolution

Step 5: Prioritize Constraints by Confidence and Impact

Not all constraints are equally important or reliable. Prioritize them into tiers.

Tier assignment criteria:

Why tier and limit: Research on LLM prompt engineering (Lakera, 2026; Palantir, 2025) shows that models follow 5-12 well-prioritized constraints reliably. Beyond ~20 constraints, adherence drops and constraints begin to cancel each other out. The tier system ensures the most important constraints get the most weight.

Total constraint budget: Aim for 15-25 total constraints across all tiers. If you have more than 25, merge overlapping constraints or demote low-confidence ones.

Priority signals:

Step 6: Build the Numeric Profile

Assemble all quantitative constraints into a structured numeric profile. This is the machine-readable half of the specification.

Numeric profile structure:

## Numeric Profile

### Readability Targets
| Metric | Target Range | Hard Floor | Hard Ceiling | Source | Confidence |
|--------|-------------|------------|--------------|--------|------------|
| FK Grade Level | [X-Y] | [min] | [max] | Readability/LD | [H/M/L] |
| Coleman-Liau Grade | [X-Y] | [min] | [max] | Readability/LD | [H/M/L] |
| Avg Sentence Length (words) | [X-Y] | [min] | [max] | Stylometric + Readability | [H/M/L] |
| Avg Word Length (chars) | [X-Y] | [min] | [max] | Readability/LD | [H/M/L] |

### Lexical Diversity Targets
| Metric | Target Range | Source | Confidence |
|--------|-------------|--------|------------|
| MTLD | [X-Y] | Readability/LD | [H/M/L] |
| MATTR-50 | [X-Y] | Readability/LD | [H/M/L] |
| Long Word % (6+ chars) | [X-Y%] | Big Five (O) + Readability | [H/M/L] |

### Function-Word Targets
| Feature | Target Rate (per 1000 words) | Stability | Source | Confidence |
|---------|------------------------------|-----------|--------|------------|
| Articles | [X-Y] | [high/mod/low] | Stylometric | [H/M/L] |
| Prepositions | [X-Y] | [high/mod/low] | Stylometric | [H/M/L] |
| 1st person singular | [X-Y] | [high/mod/low] | Stylometric + Big Five (N) | [H/M/L] |
| 1st person plural | [X-Y] | [high/mod/low] | Stylometric + Big Five (E) | [H/M/L] |
| Conjunctions | [X-Y] | [high/mod/low] | Stylometric | [H/M/L] |
| [other measured features] | ... | ... | ... | ... |

### Psycholinguistic Dimension Targets
| LIWC Dimension | Target % of Vocabulary | Priority Rank | Source | Confidence |
|---------------|----------------------|--------------|--------|------------|
| Cognitive processes | [X-Y%] | [1-N] | LIWC | [H/M/L] |
| Social processes | [X-Y%] | [1-N] | LIWC | [H/M/L] |
| Affective processes | [X-Y%] | [1-N] | LIWC | [H/M/L] |
| [other dominant dimensions] | ... | ... | ... | ... |

### Rhetorical Structure Targets
| Feature | Target Rate | Unit | Source | Confidence |
|---------|------------|------|--------|------------|
| Hedges | [X-Y] | per 100 clauses | Rhetorical Structure | [H/M/L] |
| Discourse markers | [X-Y] | per paragraph | Rhetorical Structure | [H/M/L] |
| Concessions | [X-Y] | per 1000 words | Rhetorical Structure | [H/M/L] |
| Rhetorical questions | [X-Y] | per 1000 words | Rhetorical Structure | [H/M/L] |

### Speech Act Distribution
| Act Type | Target % | Range | Source | Confidence |
|----------|---------|-------|--------|------------|
| Explaining | [X%] | [X-Y%] | Speech Act | [H/M/L] |
| Advising | [X%] | [X-Y%] | Speech Act | [H/M/L] |
| Challenging | [X%] | [X-Y%] | Speech Act | [H/M/L] |
| Supporting | [X%] | [X-Y%] | Speech Act | [H/M/L] |
| Questioning | [X%] | [X-Y%] | Speech Act | [H/M/L] |

Step 7: Write the Prose Specification

Translate the numeric profile into natural-language writing instructions. This is the human-readable half and the part most directly usable as an LLM system prompt.

Prose specification structure:

The prose specification has four sections, ordered by decreasing importance:

Section A: Voice Identity (from Persona Archetype + MDPI + Big Five composite)

• Who is this voice? What role does it play?
• What is the emotional register? (valence boundaries, arousal level)
• What is the social stance? (peer, authority, subordinate, observer)
• 2-3 sentences maximum. This is the anchor.

Section B: Core Writing Rules (Tier 1 constraints in natural language)

• The 5-8 most important constraints, stated as clear instructions
• Each rule cites its source but reads as natural prose
• These rules are non-negotiable

Section C: Stylistic Preferences (Tier 2 constraints)

• The 5-10 secondary constraints, stated as preferences
• "Prefer X over Y" or "When possible, use X"
• These shape the voice but allow flexibility

Section D: Contextual Nuance (Tier 3 + register variation + audience adaptation)

• How the voice shifts across contexts (if register variation was detected)
• Audience-awareness rules (if social graph analysis was available)
• Soft constraints that add authenticity

Writing the prose:

For each section, convert numeric constraints into natural instructions:

What NOT to translate into prose: Function-word frequencies, punctuation rates, and other unconscious features should remain in the numeric profile only. Telling an LLM "use articles at 6.5% rate" is counterproductive -- it forces conscious attention to something that should emerge naturally from following higher-level voice constraints. These numeric targets are for post-hoc validation, not real-time instruction.

Step 8: Validate Specification Implementability

Before finalizing, check that the specification can actually be followed.

Validation checks:

The exemplar test is critical. Write 2-3 test sentences following only the Tier 1 constraints. If you cannot produce sentences that satisfy all of them simultaneously, the specification is over-constrained and must be relaxed.

Step 9: Write the Report

Write all findings to docs/analysis/24-style-specification.md.

Report Output Template

The final report MUST be written to docs/analysis/24-style-specification.md with this structure:

# Style Specification

## Analysis Coverage

### Available Analyses
| # | Analysis | Report Location | Findings Used | Confidence |
|---|----------|----------------|---------------|------------|
| [N] | [name] | docs/analysis/NN-name.md | [count] findings | [overall H/M/L] |

### Coverage Assessment
- **Specification type:** [Full / Strong / Moderate / Minimal]
- **Dimensions covered:** [N] of 12
- **Dimensions missing:** [list with reasons]
- **Overall confidence:** [assessment]

## Constraint Derivation

### Extracted Findings
[Per-analysis finding tables from Step 2]

### Analysis-to-Constraint Translations
[Per-finding constraint translations from Step 3, organized by source analysis]

### Conflict Resolutions
| Conflict | Source A | Source B | Resolution | Rationale |
|----------|---------|---------|------------|-----------|
| [description] | [finding] | [finding] | [resolution] | [why] |

## Prioritized Constraints

### Tier 1: Core Constraints (MUST)
| # | Constraint | Source Analysis | Confidence | Rationale |
|---|-----------|----------------|------------|-----------|
| 1 | [constraint] | [source] | [H/M] | [why Tier 1] |

### Tier 2: Strong Preferences (SHOULD)
| # | Constraint | Source Analysis | Confidence | Rationale |
|---|-----------|----------------|------------|-----------|
| 1 | [constraint] | [source] | [M/H] | [why Tier 2] |

### Tier 3: Soft Suggestions (MAY)
| # | Constraint | Source Analysis | Confidence | Rationale |
|---|-----------|----------------|------------|-----------|
| 1 | [constraint] | [source] | [L/M] | [why Tier 3] |

### Omitted (Insufficient Evidence)
| Dimension | Reason Omitted | What Would Be Needed |
|-----------|---------------|---------------------|
| [dimension] | [reason] | [what analysis or data is missing] |

## Numeric Profile
[Complete numeric profile from Step 6]

## Prose Specification

### A. Voice Identity
[2-3 sentence voice anchor]

### B. Core Writing Rules
[Tier 1 constraints as natural prose instructions]

### C. Stylistic Preferences
[Tier 2 constraints as preference statements]

### D. Contextual Nuance
[Tier 3 + register variation + audience rules]

## Validation

### Constraint Summary
- **Total constraints:** [N]
- **Tier 1:** [N] | **Tier 2:** [N] | **Tier 3:** [N]
- **Internal consistency:** [pass/fail with notes]
- **Numeric feasibility:** [pass/fail with notes]

### Exemplar Sentences
[2-3 sentences written to satisfy all Tier 1 constraints, demonstrating the voice]

### Known Gaps
[Dimensions not covered, with impact assessment]

## NNGroup Tone Dimensions
Map the specification to the four standard tone-of-voice dimensions:
| Dimension | Position | Evidence |
|-----------|----------|----------|
| Formal <---> Casual | [position on spectrum] | [which analyses support this] |
| Serious <---> Funny | [position on spectrum] | [which analyses support this] |
| Respectful <---> Irreverent | [position on spectrum] | [which analyses support this] |
| Enthusiastic <---> Matter-of-fact | [position on spectrum] | [which analyses support this] |

## Limitations and Caveats
- Style specification is derived from analysis of [corpus description], not from direct observation of the author's intent
- Confidence levels propagate from upstream analyses; low-confidence upstream = low-confidence constraint
- The specification captures patterns in a specific corpus and time period; the author's voice may vary in other contexts
- LLMs cannot consciously control all specified features (especially function-word rates); numeric targets for unconscious features are for post-hoc validation only
- [Corpus-specific limitations]
- [Analysis-gap limitations]

## References
- [Upstream analysis reports cited]
- Nielsen Norman Group (2016). The Four Dimensions of Tone of Voice.
- Boyd, R.L., et al. (2022). LIWC-22 Manual.
- [Other methodological references used]

Good Patterns

• Map every analysis to a specific constraint using the mapping table. No analysis output should sit unused unless confidence is too low.
• Use ranges, not exact values. A finding of "FK 12.3" becomes a constraint of "FK 11-14." Exact values create false precision and are impossible for writers or LLMs to hit.
• Prioritize constraints by confidence. High-confidence findings from direct measurement (readability, stylometric) become Tier 1. Low-confidence inferences (personality traits with wide bands) become Tier 3 or are omitted.
• Resolve conflicts explicitly. When two analyses disagree, document both findings, explain the conflict, and state how it was resolved. Never silently drop a finding.
• Produce both numeric and prose forms. The numeric profile enables post-hoc validation; the prose specification enables real-time writing guidance. Both are needed.
• Keep total constraints to 15-25. Research shows diminishing returns beyond this range. Merge overlapping constraints rather than listing every sub-finding.
• Separate conscious from unconscious features. Function-word rates belong in the numeric profile for validation, not in the prose specification as instructions.
• Write exemplar sentences that demonstrate all Tier 1 constraints working together. If you cannot, the specification is over-constrained.
• Cite sources for every constraint. Traceability enables future revision when upstream analyses are re-run.
• Document gaps. Missing analyses are information; they tell the reader which constraint categories lack evidential backing.

Anti-Patterns

Boundaries

This skill SHOULD:

• Map all available analysis outputs to concrete writing constraints
• Translate findings using ranges, not exact values, scaled to confidence
• Detect and resolve conflicts between analyses using the resolution protocol
• Prioritize constraints into tiers (Core/Strong/Soft) by confidence and impact
• Produce both a numeric profile (for validation) and prose specification (for writing guidance)
• Limit total constraints to 15-25 to maintain LLM and human followability
• Document every gap where analyses were unavailable
• Write exemplar sentences demonstrating the specification
• Cite the source analysis for every constraint
• Write the complete specification to docs/analysis/24-style-specification.md

This skill should NOT:

• Invent constraints not supported by any analysis finding
• Include all findings regardless of confidence (low-confidence findings should be Tier 3 or omitted)
• Produce a specification longer than an LLM context window can reasonably process (~2000 words for the prose section)
• Skip conflict resolution when analyses disagree
• Translate unconscious features (function-word rates) into conscious prose instructions
• Re-analyze raw data (accept upstream report findings as given)
• Build a specification from fewer than 3 analyses
• Merge findings from different authors into a single-voice specification
• Claim the specification captures the author's "true voice" (it captures patterns in a specific corpus)
• Produce only a numeric profile without prose, or only prose without numeric targets

Insufficient Data Handling

When Some Analyses Were Not Completed

When Analyses Produce Contradictory Results

Follow the conflict resolution protocol in Step 4. Additional handling:

When Confidence Is Low Across the Board

If most available analyses report low confidence:

…(truncated)