RonanCodes/voice-profile

Voice Profile

Build a portable voice-profile.md that any AI can read to reproduce the user's writing voice. Based on Hassid's methodology (Jan 2026): voice is mostly refusals, not preferences, so the interview captures what the user would never write.

Source documents:

• Online: https://ruben.substack.com/p/i-am-just-a-text-file
• Vault concept page (user only): ai-research:voice-extraction-methodology

Usage

/ro:voice-profile start         # begin a new interview (or resume if in progress)
/ro:voice-profile resume        # explicit resume — pick up where you left off
/ro:voice-profile status        # show progress (questions answered per category)
/ro:voice-profile compile       # build voice-profile.md from current state
/ro:voice-profile show          # cat the compiled voice-profile.md
/ro:voice-profile path          # print canonical file paths (state + output)

The interview is ~100 questions and takes 60–120 minutes. It is meant to be run in chunks across multiple sessions. State persists between sessions.

How the interview works

Claude (you, reading this) conducts the interview live in chat. The user answers in chat. Their answers get persisted to a state file by you (Edit/Write the JSON directly).

Categories and counts

| # | Category | Questions | Probes (the area to interrogate) | |---|----------|-----------|----------------------------------| | 1 | Beliefs & Contrarian Takes | 15 | Beliefs others in the field don't share; hot takes; conventional wisdom they think is wrong | | 2 | Writing Mechanics | 20 | Sentence structures, openers/closers, punctuation/formatting habits, words they overuse / love / never use | | 3 | Aesthetic Crimes | 15 | What makes them cringe in others' writing; nails-on-chalkboard phrases; lazy content patterns | | 4 | Voice & Personality | 15 | Humor, serious vs casual, handling disagreement, sound when excited vs skeptical | | 5 | Structural Preferences | 15 | How they organise ideas, lists/headers/bullets, transitions, default content shapes | | 6 | Hard Nos | 10 | Topics they'd never write about, approaches they'd never take, lines they won't cross | | 7 | Red Flags | 10 | What makes them distrust a piece of content immediately; signals someone doesn't know what they're talking about | | Total | | 100 | |

Hassid did not publish the verbatim 100 questions. Generate them dynamically within each category, guided by the probes. This adapts to the user's prior answers and avoids overlap.

Six interview rules (load into your behavior)

1. ONE question at a time. Wait for the response before asking the next.
2. Push back on vague answers. If they say "I write conversationally," ask: "Conversational how? Give me a sentence you've written that captures it, and one that's lazy-conversational."
3. Demand specific examples. Quotes from real things they've written. Real phrases they hate. Real conventional-wisdom claims they reject.
4. Call out contradictions. If Q3 contradicts Q12, surface it: "You said X earlier, now you're saying Y — which is it?"
5. Follow interesting threads deeper. If a one-line answer hints at something rich, drill in for 1–2 follow-ups before moving on.
6. Don't accept "I don't know" easily — reframe. "If you had to pick one, which side of this do you lean?" or "What would your gut answer be even if you're not 100% sure?"

The point is rich, specific answers. A 3-word vague answer is worthless. A 4-paragraph specific answer is the goal.

Dispatch

| Arg | Behavior | |-----|----------| | start | If no state exists, run scripts/start.sh to create one (asks for user's name first), then begin Q1 of category 1. If state exists with status in_progress, behave as resume. | | resume | Read state, print summary ("You're 23 of 100 done, currently in Category 2: Writing Mechanics"), then ask the next unanswered question. | | status | Run scripts/status.sh. Print only — do not ask a question. | | compile | Run scripts/compile.sh. Compile current state into voice-profile.md. Works even if interview is incomplete (will mark unanswered sections). | | show | Run scripts/show.sh to print the compiled file. | | path | Run scripts/path.sh to print where state and output live. |

Recording answers (your job during the interview)

After each user answer that has passed the rules (specific, exampled, non-contradictory), update the state file at ~/.claude/voice-profile-state.json directly via the Edit tool. Append a new entry to the category's answered array:

{
  "q": "<the question you asked>",
  "a": "<their answer, verbatim>",
  "follow_ups": [
    {"q": "<your push-back>", "a": "<their refined answer>"}
  ],
  "answered_at": "<ISO 8601 timestamp>"
}

Then ask the next question. Do not batch — record after every settled answer so the work is never lost mid-session.

When a category finishes (e.g., 15/15 in Beliefs):

1. Acknowledge the milestone ("That's category 1 done.").
2. Summarize 2–3 things you learned about their voice from that category.
3. Ask if they want to continue or pause. If pause: stop, remind them of /ro:voice-profile resume.

When all 100 are done, run scripts/compile.sh and tell them where the file is.

First-time setup

No setup. The skill creates ~/.claude/voice-profile-state.json on start and ~/.claude/voice-profile.md on compile. Both are gitignored by default.

Vault- or project-backed state (recommended for durable / wiki-integrated use)

State and output default to ~/.claude/ but are overridable via env vars (see scripts/common.sh):

• VOICE_PROFILE_STATE — path to the JSON state file
• VOICE_PROFILE_OUTPUT — path to the compiled markdown

To keep a profile inside an llm-wiki vault (versioned, browsable in Obsidian, resumable across machines), export both before any subcommand:

export VOICE_PROFILE_STATE="$PWD/vaults/llm-wiki-voice/scratchpad/voice-profile-state.json"
export VOICE_PROFILE_OUTPUT="$PWD/vaults/llm-wiki-voice/wiki/entities/voice-profile-<name>.md"

Resume discovery (important). When the user asks in natural language to "continue / pick up my voice profile" — without typing /ro:voice-profile resume — do NOT start a fresh interview. Locate existing state first, in this order:

1. A loaded memory note about a paused voice interview (it names the state path).
2. vaults/*/scratchpad/voice-profile-state.json under the current repo.
3. The default ~/.claude/voice-profile-state.json.

If a state file with status: in_progress is found, export VOICE_PROFILE_STATE/VOICE_PROFILE_OUTPUT to point at it, run status to reload progress, then behave as resume.

Output format

voice-profile.md follows Hassid's template:

# VOICE PROFILE: <Name>
## Core Identity (2-3 sentence essence — only summary section)
## SECTION 1: BELIEFS & CONTRARIAN TAKES (Q1–Q15 with full answers)
## SECTION 2: WRITING MECHANICS (Q16–Q35)
## SECTION 3: AESTHETIC CRIMES (Q36–Q50)
## SECTION 4: VOICE & PERSONALITY (Q51–Q65)
## SECTION 5: STRUCTURAL PREFERENCES (Q66–Q80)
## SECTION 6: HARD NOS (Q81–Q90)
## SECTION 7: RED FLAGS (Q91–Q100)
## QUICK REFERENCE CARD
  - Always
  - Never
  - Signature Phrases & Structures
  - Voice Calibration (key quotes from the interview)
## HOW TO USE THIS DOCUMENT (ANTI-OVERFITTING GUIDE)
  - Frequency labels: HARD RULE / STRONG TENDENCY (70–80%) / LIGHT PREFERENCE
  - Litmus test: "Does this sound like something I would actually write, or does it sound like an AI trying very hard to imitate me?"
  - Format adaptation: tweet ≠ newsletter ≠ LinkedIn ≠ long-form
  - Spirit Over Letter
## INSTRUCTIONS FOR CLAUDE (the file is the context — every prompt should start by reading it)

The Quick Reference Card and Anti-Overfitting Guide are derived from the answers, not asked directly. Compile them after the 100 are done.

What this skill does NOT do

• Does not write the questions for you. You generate them in chat, using the category probes as your guide.
• Does not auto-resume. If state exists, the user still types start or resume — no surprise re-engagement on every session.
• Does not publish the file anywhere. It writes to ~/.claude/voice-profile.md only. The user is responsible for copying it to other AIs / hosts.
• Does not produce per-format variants (tweet voice vs LinkedIn voice vs newsletter voice). The single file documents tendencies; the consumer (e.g. /ro:write-copy or a future /draft skill) decides which tendencies apply per format.

See also

• ro:write-copy — once voice-profile.md exists, this skill should read it as additional context for any drafting work.
• reference.md — full question-bank guidance, output template, anti-overfitting layer detail.
• Original methodology: https://ruben.substack.com/p/i-am-just-a-text-file