voice-dna

Voice DNA — LinkedIn Style Extraction & Analysis

Extract a person's writing voice from their LinkedIn posts. Produce a reusable style profile with an 8-dimension analysis and a Prompt Engineering Guide for AI-assisted writing in that voice.

What This Skill Produces

1. Corpus file — all LinkedIn posts with metadata (projects/blogging/corpus/{handle}/linkedin-posts.md)
2. Mode classification — posts categorized into 6 posting modes (.work/style-research/{handle}-mode-classification.md)
3. Dimension analyses — intermediate files for dims 1-4 and 5-8 (.work/style-research/)
4. Final style profile — synthesized profile with Prompt Engineering Guide (projects/blogging/analyses/{handle}.md)
5. Blended voice profile (Phase 7) — merged voice DNA from multiple profiles (projects/blogging/analyses/blended-voice.md)
6. my-voice skill (Phase 7) — auto-generated skill for content creation (~/.claude/skills/my-voice/)

Prerequisites

• Chrome MCP tools (authenticated LinkedIn session in Chrome)
• Subagent capability (mode classification + parallel dimension analysis)
• The linkedin skill handles Chrome MCP setup; this skill handles the analysis workflow

Workflow

Phase 1: Interview

Before extracting anything, understand the user's intent. Ask:

1. Who? — Name, LinkedIn handle/URL. If not provided, search LinkedIn via Chrome.
2. Why? — This shapes the output, especially the blending notes:
   • Content creation — user wants to write like this person → focus blending notes on adoptable techniques
   • Voice blending — user wants to combine this voice with their own/others → focus on non-negotiable vs blendable elements
   • Competitor study — user wants to understand a peer's style → focus on differentiation opportunities
3. What aspects interest you most? — Humor, authenticity, structure, persuasion, etc. This guides emphasis in analysis.
4. Audience context — Who does this person write for? Who does the USER write for? The gap shapes the audience translation table in blending notes.
5. Any known context? — Role, industry, native language (affects human-sound markers analysis), relationship to the user.

If the user already provided answers in their initial request, don't re-ask — incorporate what they said.

Phase 2: Extraction

Read references/voyager-api.md for the full technical reference.

Summary of the approach:

1. Navigate to https://www.linkedin.com/in/{handle}/recent-activity/all/ in Chrome
2. Capture the Voyager API endpoint via read_network_requests after clicking "Show More"
3. Extract profileUrn, queryId, and CSRF token
4. Paginate through all posts using paginationToken (NOT start parameter)
5. Output JSON chunks via console.log() channel (JS return value gets blocked by content filter)
6. Parse into corpus markdown using scripts/parse-voyager-corpus.py

Corpus format (one post per section):

## Post {N}

**Date:** YYYY-MM-DD
**Type:** Original/Repost
**Engagement:** {likes} likes, {comments} comments
**URL:** {url}

{full post text}

---

Save to: projects/blogging/corpus/{handle}/linkedin-posts.md

Update .work/style-research/scratchpad.md with extraction stats (post count, date range, originals vs reposts).

If Voyager API fails (can't capture endpoint, auth issues), fall back to DOM scraping — see the fallback section in references/voyager-api.md. DOM scraping works better on the user's own profile than on external profiles.

Phase 3: Mode Classification

Spawn 1 subagent to classify all posts. The agent needs:

• The corpus file
• references/dimensions.md (contains posting mode definitions and classification procedure)

Agent prompt template:

You are classifying LinkedIn posts into posting modes for a voice DNA analysis.

Read these files:
1. {corpus_path} — the full post corpus
2. {skill_path}/references/dimensions.md — Part 1 has the 6 posting modes and classification procedure

For each post, assign exactly one mode: authentic_voice, quick_reshare, marketing_provided, event_promo, thought_piece, or minimal_legacy.

Focus on the most recent 500 posts if the corpus is larger. For older posts, do temporal sampling (every 5th post) to confirm distribution holds.

Output format — save to {output_path}:
1. Summary table (mode, count, percentage)
2. Per-post classification: Post number, mode, confidence (high/medium/low), brief rationale for non-obvious classifications
3. Key observations about corpus composition

Subject-specific notes: {any context from the interview about this person's role, platform, likely mode distribution}

After classification completes: Present the summary table to the user for sanity-check before proceeding. Let them flag any misclassifications or patterns you missed.

Phase 4: 8-Dimension Analysis

After user approves mode classification, spawn 2 parallel subagents:

Agent A — Dimensions 1-4 (Voice, Language, Humor, Thought Leadership):

You are analyzing the writing style of {name} across 4 dimensions.

Read these files:
1. {corpus_path} — full corpus
2. {mode_classification_path} — use ONLY authentic_voice and thought_piece posts
3. {skill_path}/references/dimensions.md — Part 2, dimensions 1-4

For each dimension:
- Position the author on every spectrum (percentage + characterization)
- Cite specific evidence (post numbers + verbatim quotes)
- Flag any anti-patterns detected
- Note temporal evolution if patterns shift across years

Special attention to: {aspects the user highlighted in interview}
Context: {person's role, native language, audience — affects language register analysis}

Save to: {output_path}

Agent B — Dimensions 5-8 (Authenticity, Structure, Persuasion, Topic Framing):

You are analyzing the writing style of {name} across 4 dimensions.

Read these files:
1. {corpus_path} — full corpus
2. {mode_classification_path} — use ONLY authentic_voice and thought_piece posts
3. {skill_path}/references/dimensions.md — Part 2, dimensions 5-8

For each dimension:
- Analyze with specific evidence (post numbers + verbatim quotes)
- Identify signature structural patterns and recurring techniques
- Note temporal evolution if patterns shift across years
- Catalog hook patterns with frequencies

Special attention to: {aspects the user highlighted in interview}
Context: {person's role, audience — affects persuasion and framing analysis}

Save to: {output_path}

Phase 5: Synthesis

After both analysis agents complete, synthesize into the final profile. Read references/output-template.md for the full template.

Either do this in the main thread (if context allows) or spawn a synthesis agent that reads:

• The corpus (for Best-Of selection and Vocabulary Fingerprint)
• Both dimension analyses
• The mode classification
• references/output-template.md (the template to follow)
• Interview context (why the user wants this, audience gap for blending notes)

The synthesis must produce ALL sections in the template, including the full Prompt Engineering Guide with:

• Voice DNA compact prompt block
• Banned words/patterns
• Encouraged patterns
• Content type templates
• Calibration examples (3-7 gold standard posts)
• Blending notes with audience translation table

Save to: projects/blogging/analyses/{handle}.md

Phase 6: Verify & Present

Before presenting to the user, verify:

• All 8 dimensions analyzed with concrete examples (post numbers + quotes)
• Mode classification covers the full corpus with distribution table
• Vocabulary Fingerprint has characteristic words, avoided words, AI-smell checklist
• Best-Of Posts has 3-7 exemplars with justification
• Prompt Engineering Guide has all 6 subsections (voice DNA, banned, encouraged, templates, calibration, blending)
• Blending notes address the audience gap identified in interview
• Temporal evolution section present (if corpus spans 2+ years)

Present to the user: a brief summary of key findings, the voice essence in 1-2 sentences, and the most surprising discovery. Point them to the full profile path.

Phase 7: Voice Blending (optional)

When the user has 2+ completed profiles and wants to synthesize a blended voice for their own content creation. This phase reads the individual profiles' Prompt Engineering Guides + Blending Notes and produces a unified voice.

When to trigger: User says "blend voices", "combine profiles", "create my voice", "synthesize voice", "mix these voices", or has completed multiple profiles and wants to create content in a blended style. Also triggers when the user asks to "write a LinkedIn post" and has analyzed multiple voices but no my-voice skill exists yet.

Step 1: Interview

If not already clear from context, ask:

1. Which profiles? — List the available profiles in projects/blogging/analyses/. Confirm which ones to include.
2. Which is the base? — The user's own profile is the identity foundation. Other profiles are influences.
3. Proportions? — Suggest a starting point (e.g., 60% base, 25% strongest influence, 15% secondary). User adjusts.
4. Target channels? — Where will this voice be used? (LinkedIn, blog, etc.)
5. Any techniques to emphasize or suppress? — User may want more of Giovanni's humor but less of Robin's profanity, etc.

Step 2: Read source profiles

For each profile, load these sections:

• Prompt Engineering Guide (Voice DNA block, Banned, Encouraged, Templates, Calibration, Blending notes)
• Signature Moves
• Key Takeaways for Style Synthesis
• Summary (for voice essence)

The blending notes in each influence profile already contain: non-negotiable elements, blendable elements, audience translation table, and what NOT to blend. These are the primary inputs.

Step 3: Synthesize (1 subagent)

Spawn a synthesis agent that reads all source profiles + references/blending-template.md and produces:

1. Blended voice profile at projects/blogging/analyses/blended-voice.md — follows the blending template
2. my-voice skill at ~/.claude/skills/my-voice/SKILL.md — auto-generated skill containing the blended voice DNA, ready to trigger on LinkedIn/blog content creation
3. Calibration posts at ~/.claude/skills/my-voice/references/calibration-posts.md — 5-7 gold standard posts (full text) from across all source corpora

The blended voice must read as one coherent person, not a patchwork. The base voice's identity (L1 markers, core positioning, audience relationship) is non-negotiable. Influence voices contribute techniques, patterns, and range — not identity.

Agent prompt template:

You are synthesizing a blended writing voice from multiple analyzed profiles.

Read these files:
1. {profile_paths} — all source profiles (focus on: Summary, Signature Moves, Key Takeaways, Prompt Engineering Guide sections)
2. {skill_path}/references/blending-template.md — the template for the blended output
3. {corpus_paths} — source corpora (for selecting calibration posts with full text)

Interview context:
- Base voice: {base_name} at {base_pct}%
- Influences: {influence_name} at {pct}%, {influence_name} at {pct}%
- Target: {channels} for {audience}
- Emphasis: {techniques to emphasize}
- Suppress: {techniques to suppress}

Produce THREE files:

1. Blended voice profile → save to {blended_profile_path}
   Follow the blending template exactly. The Voice DNA block must be a single coherent prompt, not three blocks merged.

2. my-voice skill SKILL.md → save to {skill_path}
   Frontmatter: name "my-voice", description triggering on LinkedIn/blog content.
   Body: Voice DNA block, Blend Origins table, Banned/Encouraged patterns, Content Type Templates, Writing Workflow, Self-Test Checklist, source profile pointers.

3. Calibration posts → save to {calibration_path}
   5-7 posts from across source corpora. Include full post text. Choose posts that best represent the TARGET blended voice (not the most extreme examples of each individual voice).

Step 4: Verify & present

Verify:

• Voice DNA reads as one person, not a patchwork
• Banned patterns are the union of all profiles
• Encouraged patterns have source attribution
• Calibration posts span all source corpora
• Self-test checklist covers AI-smell detection
• my-voice skill frontmatter will trigger correctly
• Content type templates are blended, not just the base voice's templates

Present to the user: the blended Voice DNA block, the dial settings table, and the self-test checklist. Point them to both files (blended profile + skill).

File Conventions

File	Path
Corpus	projects/blogging/corpus/{handle}/linkedin-posts.md
Mode classification	.work/style-research/{handle}-mode-classification.md
Dims 1-4 analysis	.work/style-research/{handle}-analysis-voice-language.md
Dims 5-8 analysis	.work/style-research/{handle}-analysis-structure-persuasion.md
Final profile	projects/blogging/analyses/{handle}.md
Progress tracking	.work/style-research/scratchpad.md (append)
Blended voice profile	projects/blogging/analyses/blended-voice.md
my-voice skill	~/.claude/skills/my-voice/SKILL.md
Calibration posts	A user-maintained calibration-posts.md (e.g. in a local my-voice/ skill or project notes)
Blending template	references/blending-template.md (this skill's reference)

{handle} = LinkedIn username slug. Derive from the profile URL.

Existing Profiles (for reference)

If you have prior analyses on disk (e.g. in projects/blogging/analyses/<handle>.md), agents can read them to calibrate output depth and format. The expected shape is one markdown file per analyzed profile, named by handle, containing the 8-dimension style profile this skill produces.

Troubleshooting

Voyager API returns same page repeatedly: You're using start for pagination instead of paginationToken. See references/voyager-api.md.

Timestamps 40 years in future: You added Twitter epoch offset. LinkedIn snowflakes use Unix epoch directly. See references/voyager-api.md.

JS return value blocked: Use console.log() + read_console_messages channel. Content filter blocks combined social media text in return values.

DOM shows only ~5 posts for external profile: Expected — DOM virtualization. Switch to Voyager API method.

Corpus too large for a single console.log: Output in chunks, read all at once, parse with scripts/parse-voyager-corpus.py.