Voice Dna Kit

Voice DNA Kit

Capture how a person or organization demonstrably writes. Produce a compact operating guide backed by observed samples instead of aspirational adjectives.

Boundary

Use this skill to preserve an existing voice.

• To design how an organization should sound, use brand-voice.
• To produce a deeper quantitative report, use style-forensics.
• To draft new content from a completed profile, use style-writer.

This skill captures and packages a voice. It does not create a new brand identity or write the final long-form content.

Inputs

Ask for:

1. Three to five representative samples from the same person or organization.
2. The contexts where the voice will be used.
3. Known phrases, structures, punctuation, or tones to avoid.
4. Any before-and-after edits that show what the owner changed.
5. One representative sample that can be reserved as a holdout, when the corpus is large enough.
6. The requested output location.

Use sample-selection-guide.md to assess the corpus. Prefer writing created or substantially edited by the voice owner. State when AI-assisted or mixed-author samples weaken confidence.

Deliverables

Create these files when the user requests a persisted kit:

• voice-dna.md with short, operational rules
• style-dna.md with measurements and supporting examples
• banned-phrases.md with generic and source-specific bans
• voice-review-checklist.md with final checks
• voice-verification.md with the comparison test and remaining mismatches
• corpus-manifest.md when provenance, collection, deduplication, exclusions, or holdout selection need a durable record

Place them in the user-selected directory. Do not assume a runtime-specific folder.

Workflow

1. Select and label the corpus

Record the author, channel, audience, approximate date, provenance, approval status, and selection reason for each sample. Define the selection rule before collecting more material. Record discovered, retained, excluded, duplicate, and failed items when collection is involved.

Exclude quoted third-party text, boilerplate, transcripts the owner did not edit, and samples from a different voice or register. Keep provenance separate from prose so URLs, titles, and wrappers do not distort measurements.

If the corpus contains distinct registers, label them. Preserve one stable core and document the controlled differences. Reserve a representative holdout before extracting rules when enough material exists.

2. Extract clean prose

Remove navigation, metadata, signatures, code, URLs, quoted material, and repeated boilerplate. Keep paragraph boundaries and deliberate formatting where they carry style.

For exact duplicate detection, normalize a comparison copy without changing the analysis copy: normalize line endings, trim surrounding whitespace, and collapse repeated whitespace. Hash the normalized text and remove exact duplicates, including duplicated parent and child pages. Record what was removed. Do not use fuzzy deduplication to discard merely similar writing without review.

3. Measure before interpreting

Resolve analyze-style.sh relative to this file and run it against the cleaned samples. Save its output as the initial style-dna.md.

Review the measurements manually. Sentence splitting and rhetorical-device detection are imperfect, so correct obvious parsing errors and attach examples to every important conclusion.

Measure at minimum:

• sentence and paragraph length
• short-sentence ratio
• punctuation frequency
• contractions and pronoun use
• hedging and confidence
• recurring openers, transitions, fragments, and landings
• vocabulary specificity and repeated terms

4. Extract operational rules

Use voice-dna-template.md. Convert stable evidence into rules that can pass or fail review.

Separate:

• stable voice traits
• context-dependent tone dials
• measured tendencies
• strict bans
• uncertain observations that need more samples

Do not turn every observed quirk into a rule. Preserve patterns that recur or clearly carry identity.

5. Build the ban list

Start with banned-ai-phrases.md, but keep only defaults relevant to the intended context. Add source-specific bans from rejected drafts, absent punctuation patterns, and repeated owner edits.

Concrete bans beat instructions such as "sound human."

6. Build the review checklist

Create checks for:

• banned patterns
• sentence and paragraph rhythm
• punctuation
• confidence and hedging
• vocabulary and specificity
• register selection
• factual and attribution integrity

7. Verify the kit

Run voice-match-test.md. Compare a baseline draft with a profile-guided draft against the reserved holdout, or against representative source samples when no defensible holdout is available.

Do not use the holdout to create or patch rules before the first comparison. Record where the guided draft moves closer, stays unchanged, or overfits. Distance or resemblance is a diagnostic, not proof of quality, truth, or authorship.

Patch the profile when a mismatch cannot be explained by an existing rule, then rerun once. Do not copy distinctive phrases from source samples merely to improve resemblance.

Quality Gate

The kit is ready when:

• every strong rule cites repeated evidence or a clear edit preference
• actual numbers or explicit unknowns support the style profile
• stable voice and contextual tone are separate
• bans are concrete and reviewable
• exact duplicates and exclusions are recorded when a manifest is warranted
• holdout validation is used when the corpus supports it, or its absence is disclosed
• the guided test is closer to the samples without copying them
• the short guide is concise enough to load routinely
• corpus limitations are disclosed