Skip to content

Skill · voiceprint · Voiceprint

voice-interview

**Mode skill.** Default-enabled on the Voiceprint specialist.

What it is

**Mode skill.** Default-enabled on the Voiceprint specialist. Runs in the Full path; skipped in the Lite and Refresh paths.

A skill is a written procedure an agent loads when a job calls for it. This one is a single file, SKILL.md, and the whole file is on this page.

Part ofVoiceprint
File
SKILL.md
Length
72 lines · 3 min read
Category
voiceprint
License
Apache-2.0
Author
wayland
Words
723

Paste it into any agent’s instructions (CLAUDE.md, AGENTS.md or a custom GPT), or add the team to Brainwrite and it arrives switched on.

Use it

Use this skill in Brainwrite.

  1. 1

    Add the Voiceprint

    Its agents carry this skill. You preview every agent and skill first; skills arrive switched on and routines paused.

    Install in Brainwrite →
  2. 2

    Brief your Chief of Staff

    “Use the voice-interview skill for this: [describe the job]. Show me the plan first, and send, post or change nothing.”

  3. 3

    Approve what matters

    The agent follows the skill and brings the result back. Anything that sends, posts or changes data waits for your approval in Ask mode.

    How approvals work →

Mode skill. Default-enabled on the Voiceprint specialist. Runs in the Full path; skipped in the Lite and Refresh paths.

When to use

Use after samples are gathered, before Compile. Goal: the meta-layer samples cannot show — who the user writes for, what they refuse to write, who they reject as a stylistic model, what they wish they did more of. Skip in the Lite path; a one-paragraph audience note replaces the interview.

Trigger phrases that should activate this mode:

  • "Build my voice profile."
  • "Let's do the full version."
  • "I have samples — what's next?"
  • "Run the 22 questions."

Procedure

1. Confirm samples are in hand. Before question one, confirm 5–10 samples (and ideally 1–2 transcripts) are in the chat. If not, route back to gather. Interview-without-samples is the failure mode this method prevents.

2. Ask one question at a time. No batching. After each answer, pick one action.

  • Accept and advance. Specific, concrete, names a real person or move. Next question.
  • Push back once. Vague ("I write conversationally"), one-word, aspirational, or self-contradicting. Ask a sharper follow-up: "Conversational how? Give me a sentence you wrote last week that felt like you." Once per question — twice becomes interrogation.
  • Note and advance. "I don't know" after a real attempt. Absence is data.

3. Run the 22 questions in six clusters. Track the cluster so the user sees the shape.

  • Audience & Purpose (Q1–3): who they write for (one real person if possible), what they want the reader to feel or do, the one thing they want to be known for saying.
  • Voice & Tone (Q4–8): how a friend would ID their writing without a byline, casual-to-formal range and the line they won't cross, swearing rules, two or three tonal adjectives that fit, words that make them cringe in other writers.
  • Refusals & Pet Peeves (Q9–11): topics they will never write, advice they refuse to give, what generic AI writing gets wrong.
  • Style & Structure (Q12–15): sentence length, paragraph shape, list-vs-prose preference, openers and closers, when headers and bullets help versus clutter, how much they explain.
  • Influences & Anti-Influences (Q16–18): two or three writers whose voice they admire (and what specifically), one or two they actively reject and why, the soundtrack-genre metaphor for their writing.
  • Subject & Stance (Q19–22): the three to five topics they write about, a take most peers disagree with, what they wish more people understood, the difference between writing they're proud of and writing that just got the job done.

4. Close cleanly. At Q22, say "done" and stop. No mid-interview summary, no thematic synthesis, no "based on your answers I'm noticing…" Analysis happens in Compile with samples in hand.

Decision rules

  • One question at a time, always. Batching feels efficient and destroys signal. Rapid-fire flattens the surfacing work.
  • Push back once, not twice. First push edits; second push performs.
  • Capture verbatim. Keep the user's exact words. The Compiler needs them.
  • Vague is data. "I don't know who I write for" is a real answer. Note it; don't invent.
  • Self-contradiction is gold. "I never hedge — but I always say 'I think' first." Compile will surface it.

Anti-patterns

  • Summarizing as you go. Premature pattern-matching anchors Compile to your interview-time guess, not the samples.
  • Accepting "engaging," "authentic," "conversational," or "professional but approachable" without a follow-up. These words name nothing. Push back every time.
  • Leading the witness. "Would you say you write like a Substack writer?" Ask open questions; tolerate silence.
  • Running the interview without samples. You'll produce an aspirational profile that contradicts the actual writing.
  • Reordering the clusters. Audience first (concrete), tone next (reflective), then refusals, style, influences, stance. Reordering breaks the warm-up.

Before / after

Brief: "Q5 — How casual or formal are you? Where's the line you don't cross?"

Before (accepting vague):

User: "I'm conversational but professional." → Voiceprint: "Got it. Next question."

After (one push-back):

User: "I'm conversational but professional." → Voiceprint: "That sentence applies to half the writers on the internet. Give me a phrase you wrote in the last week that felt casual, and one that felt as formal as you go. Then name the line."

The first version produces a generic profile. The second produces calibration data.

Put this skill to work.

Download Brainwrite, add the team that carries it, and tell your Chief of Staff what needs doing.

macOS today. Windows and Linux are coming soon.