brand-voice

Brand Voice

Adapted from ECC's brand-voice (MIT, (c) Affaan Mustafa). See LICENSE.

Build a durable [VOICE PROFILE] from real source material -- real messages that got replies, real openers that landed meetings -- then reuse that profile across every drafting skill instead of re-deriving style or defaulting to generic AI sales copy.

This skill owns the VOICE PROFILE format. Every drafting skill that needs voice consistency cites [VOICE PROFILE] and reads the output this skill produces. Defining a competing voice structure anywhere else is an anti-pattern.

Governing rule: rules/common/selling-principles.md -- voice is a style overlay, never a license to fabricate claims or misrepresent the sender. A profile captures how to say things; what to say is governed by product-knowledge and the deal context.

When to Activate

Activate this skill when:

• A rep wants drafts that sound like them, not like a template.
• A cold-outreach or follow-up skill is producing copy the rep won't send.
• Onboarding a new rep: capturing their writing style before they have AI assistance in their workflow.
• Moving to a new channel (LinkedIn DM vs. email vs. cold-call opener) and the per-channel voice notes need to be established or updated.
• A manager wants a consistent team voice for a new segment play.
• The existing profile is stale (>90 days) and a refresh sample exists.

Do not activate to make generic copy "sound less generic" by adjusting one or two surface words. If the problem is that a draft is too long or too formal, that is a revision task, not a voice-profile task. If the draft is fabricating claims, fix the claim source -- the voice layer cannot fix that.

Source Priority

Use the strongest real source set available, in this order:

1. Sent emails that got a positive reply (booked meeting, continued thread, warm response). These are the highest signal -- they worked.
2. Cold call openers and voicemail scripts the rep actually uses.
3. LinkedIn connection requests or DM sequences the rep wrote themselves.
4. Meeting recap or follow-up notes the rep sends after discovery calls.
5. If the rep has no usable sample: manager-supplied exemplar emails from a teammate whose style fits the segment.

Do not use generic sales email templates, training library examples, or any content the rep did not author or explicitly endorse as "my voice". The profile is a fingerprint, not a persona assigned from the outside.

Minimum sample: 5 messages. Ideal: 10-20. More than 20 adds diminishing returns unless the rep writes in clearly different modes for different segments.

Collection Workflow

1. Gather the source set. Ask the rep to paste or attach 5-20 real sent messages. If pulling from an email export, use only the rep's own sent side of threads -- not the prospect's replies.
2. Separate by mode if needed. If the sample clearly splits into "first touch cold" vs. "warm follow-up" vs. "post-demo check-in", note the split in the profile rather than averaging them into a single mush.
3. Extract the signal fields below. Work field by field; do not summarize the whole batch into one vague paragraph.
4. Produce the [VOICE PROFILE] block using the canonical format defined in this skill (see "Voice Profile Format" section below).
5. Confirm with the rep. Read back the Banned Moves and Preferred Moves lists specifically -- those are the easiest to get wrong and the most important to get right.

What to Extract

For each source sample, note the following signal dimensions:

• Sentence length and rhythm: Short and punchy? Long and explanatory? Mixed? Does it fragment intentionally?
• Compression: Dense (every sentence adds new information) or explanatory (context given before claims)?
• Opening move: Does the rep open with the prospect ("You just...")? With themselves ("I'm...")? With a question? With a fact?
• Personalization style: Research-led specificity, or pattern-matched persona assumption? How many lines before the rep gets to the point?
• Question use: How many questions per message? Are they rhetorical, direct, or hypothetical? Does the rep use a question to close?
• Claim style: Are claims backed with a number or reference immediately, or stated and left to imply? How sharply are comparisons drawn?
• CTA pattern: Single ask or multiple? Soft ("worth a chat?") or direct ("15 min Thursday 2pm or Friday 10am -- which works?")? Does the rep offer a low-commitment step first?
• Sign-off: Formal (Best regards), minimal (one word + name), or dropped entirely?
• Channel-specific norms: What the rep does differently in LinkedIn vs. email vs. voicemail.

Voice Profile Format

Produce a [VOICE PROFILE] block using this exact structure. Keep bullets concrete -- short phrases, not essays. Every Banned Move must be observable in the source set or explicitly requested; do not add bans speculatively.

[VOICE PROFILE]
===============
Rep:
Segment focus:
Confidence: high / medium / low (based on sample size and consistency)
Sample count:
Last updated:

Source Set
- <description of source 1, e.g. "6 cold emails, SaaS AE targets, Q1 2026">
- <description of source 2>
- <note any mode splits, e.g. "cold vs. warm voice differ -- see Channel Notes">

Rhythm
- <sentence length and pacing: e.g. "short sentences, 1-2 per paragraph, no filler">

Compression
- <density: e.g. "high -- one idea per sentence, no lead-in context before the claim">

Opening Move
- <how the rep opens: e.g. "prospect-first: names the trigger before anything else">

Personalization Style
- <e.g. "one concrete research line max, then the pivot; does not pad with flattery">

Question Use
- <frequency and type: e.g. "one question per message max, always at the end, direct ask">

Claim Style
- <e.g. "number or reference within the same sentence as the claim; no bare adjectives">

Preferred Moves
- <move 1: e.g. "names the trigger in line one">
- <move 2>
- <move 3>

Banned Moves
- <ban 1: e.g. "no 'I hope this finds you well' or any variant">
- <ban 2>

CTA Pattern
- <e.g. "single ask, specific times offered, 'worth a chat?' as the minimal variant">

Sign-off
- <e.g. "first name only, no 'Best' or 'Thanks'">

Channel Notes
- Email: <what changes in email vs. baseline>
- LinkedIn: <e.g. "shorter; no formal sign-off; opener is one line max">
- Cold call opener: <e.g. "lead with the trigger; permission ask within first 10 seconds">

Hard Bans (Sales-Context Defaults)

These are banned regardless of the rep's personal sample, because they are statistically low-reply patterns. A rep's profile can override any item in this list only if the rep's source set shows they work in their specific book of business and the rep explicitly confirms them.

• "I hope this email finds you well" or any variant ("Hope you're doing great", "Trust all is well with you").
• "I wanted to reach out" -- start with the reason, not the meta-statement.
• "Just following up" as an opener -- say what changed or why now.
• Manufactured flattery ("Love what you're doing at [Company]" with no specific detail).
• The feature-dump opener (listing three things the product does before saying anything about the prospect).
• Passive CTA ("Let me know if you're interested" / "Feel free to reach out").
• Stacked asks (more than one request in a single message).
• "Excited to connect" or "Would love to chat" without a concrete agenda.
• All-caps emphasis used as a substitute for a sharp claim.
• Generic P.S. lines ("P.S. Here is a case study you might find relevant").

Persistence Rules

• Reuse the latest confirmed [VOICE PROFILE] across related tasks in the same session without re-deriving it.
• If a rep requests a durable artifact, save the profile to .claude/escc/voice/<rep-slug>.md (workspace-local; never committed with personal data unless the rep explicitly requests repo tracking).
• For a durable per-account style overlay (how a specific account writes, to layer on top of the rep base profile), see "Per-Account Voice Overlay" below — stored at .claude/escc/voice/account/<account>.md.
• If the drafting skill is cold-outreach, outbound-sequences, follow-up-ops, reply-handling, email-outbound-ops, meeting-followthrough, or inbox-triage, check whether a session-local [VOICE PROFILE] already exists before starting a new collection.
• Do not store voice profiles in HubSpot records. They are a local style artifact, not a CRM field.

Per-Account Voice Overlay

A [VOICE PROFILE] captures how the rep writes. A per-account voice overlay captures how a specific account writes, so a draft to that account mirrors their register and vocabulary on top of the rep's base profile.

Contract for consumers. Every voice-consuming skill (cold-outreach, outbound-sequences, follow-up-ops, reply-handling, email-outbound-ops, meeting-followthrough, inbox-triage) layers this overlay on the rep base profile whenever it drafts to a known account with prior correspondence: load it with escc voice show "<account>". The overlay only adjusts register and word choice — the rep base profile still wins on Banned/Preferred Moves, and facts still come only from approved product-knowledge.

• Storage: .claude/escc/voice/account/<account>.md (gitignored — it is mined from real correspondence and never belongs in the source repo). It layers on the rep base profile at .claude/escc/voice/<rep-slug>.md; it never replaces it.
• A draft is: rep base voice × buyer-role register × this-account register × the account's mirrored lexicon. The base profile still wins on the rep's own Banned Moves and Preferred Moves; the overlay only nudges register (formality, sentence length, question rate, greeting/sign-off) and word choice toward this account.
• Build or refresh it deterministically with escc voice account "<account>" --input '{"texts":[...]}', then read it back with escc voice show "<account>". The extractor is no-ML and deterministic — it does not "interpret" the account, it tallies observable style.
• Buyer side only, via quarantine. The texts are the buyer's words — the emails they sent you and their turns in a call transcript — gathered through the read-only quarantine/thread path (transcript-analyzer, email-outbound-ops), never the rep's own sent copy. Raw bytes never reach a privileged context.

The style/content split is enforced here, not just stated. The overlay mirrors the buyer's words, never their claims or numbers. The lexicon is pure-alphabetic terms only — a metric, a percentage, or a currency figure can never become a term, and a source sentence is never echoed into the overlay. You may sound like the account; you may never repeat their figure back as if it were our proof. Facts and metrics come only from approved product-knowledge. (Enforced by scripts/lib/account-register.js + scripts/lib/voice-overlay.js, pinned by tests/unit/content-guard-lexicon-leak.test.js.)

Examples

Profile built from 8 cold emails:

[VOICE PROFILE]
===============
Rep: A. Patel
Segment focus: SaaS mid-market RevOps
Confidence: high
Sample count: 8 sent emails with positive reply
Last updated: 2026-06-16

Source Set
- 8 cold emails to RevOps Directors, sent Q4 2025 - Q1 2026, all had positive reply
  or booked meeting

Rhythm
- 1-2 sentences per paragraph; no paragraph over 3 lines; deliberately fragmented
  for scannability

Compression
- high; no contextual preamble before the claim; no "as you may know" framing

Opening Move
- always leads with a specific company or persona observation within the first line
  (trigger, role change, or segment-specific assumption)

Personalization Style
- one research line tied to a real signal (job post, announcement, funding); moves
  to the pivot immediately; never pads

Question Use
- one per message, at the end; always a direct yes/no or specific time ask

Claim Style
- cites a reference or number in the same sentence; never bare adjectives
  ("fast" or "better"); always "X faster" or "teams like yours report Y"

Preferred Moves
- Opens with prospect's situation, not own product
- Lands the value in one short sentence before the CTA
- Uses one concrete proof point from product-knowledge, attributed correctly

Banned Moves
- "I hope this finds you well" or any greeting filler
- "I wanted to reach out" as an opener
- Feature-dump before the prospect hook
- Stacked asks

CTA Pattern
- single ask; offers two specific time slots; "worth a 15?" as the minimum variant

Sign-off
- first name only

Channel Notes
- Email: baseline above
- LinkedIn: opener compressed to one line; no sign-off; uses "open to it?" as CTA
- Cold call opener: trigger first ("I saw you just posted for a Revenue Ops
  Director -- that usually means..."); permission ask by second sentence

When a drafting skill should cite the profile:

cold-outreach: building first-touch email for A. Patel -> GlobalBank RevOps VP
brand-voice: [VOICE PROFILE] loaded (A. Patel, confirmed 2026-06-16)
cold-outreach: applying profile --
  - opener: GlobalBank just raised Series C; leading with funding trigger
  - compression: high; claim in line two
  - one proof point from product-knowledge PP-031 (onboarding velocity)
  - CTA: "worth a 15 this week?" -- no stacked ask, specific times in follow-up
  - sign-off: "A." only

Anti-patterns

• Ignoring the source set and generating a profile from scratch. A profile must be sourced from real material. If no sample exists, say so and ask for one; do not invent a representative style.
• Averaging a split voice into one profile. If the rep writes very differently for cold vs. warm, document two modes; do not blend them.
• Adding bans speculatively. Only ban what is observable in the source set or what the rep explicitly requests. Unsourced bans create a profile that does not actually reflect the rep.
• Using the voice profile to excuse fabricated claims. Voice is how you say it. What you say is governed by product-knowledge. The profile does not override the evidence-first requirement.
• Storing the profile in HubSpot or in a shared CRM field. Voice profiles are personal writing fingerprints; they belong in the local workspace, not in the system of record.
• Re-deriving style on every draft. Build the profile once, then cite it. Do not produce slightly different voice analysis on each message -- that defeats the purpose of a durable profile.
• Applying the profile to a different rep's outreach without confirmation. A profile is rep-specific. Do not reuse A. Patel's profile for K. Lee's messages without explicit approval and confirmation.

Related

• Feeds: cold-outreach, outbound-sequences, follow-up-ops, reply-handling, email-outbound-ops, meeting-followthrough, inbox-triage, meeting-booking (opener scripts).
• Proof for claims inside voiced messages comes from product-knowledge.
• Per-account context (prospect name, role, trigger details) comes from account-memory.
• Outbound compliance constraints (opt-out, cadence limits) are governed by rules/common/outbound-compliance.md and are not overridden by voice preferences.
• This skill is command-less (auto-trigger / sub-workflow). Drafting skills invoke it; it is not a standalone slash command.