| name | voice-sample |
| description | Voice sampling — study how Peter actually talks in conversations (not published pages). Captures rhythm, openings, vocabulary, humor. Accumulates observations over sessions. |
| user_invocable | true |
Voice Sample — Atomic Skill
Purpose
Before writing copy, you need to know how Peter actually sounds. Not how his published pages sound — how HE sounds when he's explaining something in conversation. This skill searches his conversation history, extracts voice patterns, and accumulates observations over time.
The voice sample is a tuning fork, not a style guide. After writing a draft, read it back and ask: does this sound like the person in those conversations? Not a polished version of him. Him.
Usage
/voice-sample [topic] # Sample how Peter talks about a specific topic
/voice-sample # General voice sampling (broader search)
Context
Read
- Conversation exports: Claude Code JSONL, ChatGPT, Claude.ai, Gemini
- Filter for Peter's messages only (not AI responses)
- Previous observations:
memory/voice-sample.md
Invoke (as needed)
N/A — this skill observes, it doesn't evaluate against other skills.
Lens Criteria (embedded)
N/A — this is an observation tool. It accumulates patterns, not verdicts.
Sources (Peter's messages only)
Filter for Peter's words, not AI responses. Search in this order:
1. Current conversation
How Peter is talking to you right now. This is the freshest, most natural sample.
2. Claude Code sessions
Location: /home/peter/.claude/projects/-home-peter-homelab-projects-active-petersalvato-com/*.jsonl
Filter: Look for "role":"human" entries.
grep -l "[topic]" /home/peter/.claude/projects/-home-peter-homelab-projects-active-petersalvato-com/*.jsonl
Then extract Peter's messages from matching sessions.
3. PKB search
/pkb search "[topic]"
The PKB indexes 52,000+ documents across all sources. If available, use it for broad searches.
4. ChatGPT exports
Location: /home/peter/homelab/knowledge/exports/JSON/conversations.json
Filter for Peter's messages (the human turns, not the assistant turns).
5. Claude.ai / Gemini exports
Location: /home/peter/homelab/knowledge/exports/consolidated_exports/consolidated_clean/
These are markdown — harder to filter speaker, but still useful for patterns.
What to Capture
From conversations, NOT from published pages:
- Sentence rhythm — Short bursts or long flowing thoughts? How does he connect ideas?
- Opening moves — How does he start explaining something? ("I think...", "so basically...", "the thing is...", "yeah let's...")
- Vocabulary — What words does he reach for naturally? What words does he never use?
- Transitions — Connectors ("and", "so", "but") or hard breaks?
- How he describes problems — Names the feeling first or the situation?
- How he describes solutions — Explains the mechanism or the result?
- Humor and tone — Dry? Self-deprecating? Matter-of-fact?
- Phrasing patterns — Recurring structures, pet phrases, how he qualifies things
- How he gives direction — Does he specify exactly or sketch loosely? Does he think out loud?
Output
Print observations in conversation. Also accumulate in the persistent memory file.
In conversation:
# Voice Sample — [topic or "General"]
## Rhythm
[How he strings sentences together]
## Opening Moves
[How he starts explaining — with examples quoted from conversations]
## Vocabulary
- Reaches for: [words he uses naturally]
- Never uses: [words absent from his speech]
## Tone
[Dry/warm/matter-of-fact/self-deprecating — with examples]
## Patterns
[Recurring structures, pet phrases, qualifiers]
## Example Quotes
[3-5 actual quotes from conversations that capture his voice]
Persistent file:
Append observations to:
/home/peter/.claude/projects/-home-peter-homelab-projects-active-petersalvato-com/memory/voice-sample.md
Each session that does copy work should ADD to this file, not replace it. Over time it becomes a detailed fingerprint. Date-stamp each addition.
Format for persistent entries:
## [DATE] — [topic or context]
- [observation]
- [observation]
- Quote: "[actual quote]"
Relationship to Voice Protocol
The Millman x Craftsman protocol (in voice/copywriting-protocol.md) defines the register: personal stakes through action, material vocabulary, no hype.
The voice sample defines the person inside that register. Two different writers could follow the same protocol and produce different copy. The voice sample ensures the copy sounds like Peter specifically.
Called By
/copy edit — Phase 1, Step 2 (while searching ideation history)/copy write — Step 3- Standalone — usable anytime to build the fingerprint
