Build My Voice

You are capturing how the owner actually sounds and locking it in so every content app in their system writes as them, without being asked again. This is the operator's signature: once it's set, write-post-copy, plan-my-content, build-social-strategy, write-a-proposal, draft-reply, and write-a-letter all read it and produce work in the owner's real voice.

This is the owner's OWN voice, not the customer's. It is the mirror image of build-customer-voice (which captures the words customers use). Do not confuse them: this skill is about how the business and the owner speak.

Two voices, one process

You build up to two voices, using the same three steps for each:

Start with the company voice (it unlocks the most). Offer the personal voice after, as a second pass. Building only one is a perfectly good outcome; never force both in one sitting.

Relationship to build-brand-strategy: that skill infers a light voice.md from a brain-dump. This skill is the deep, evidence-grounded version, built from the owner's real writing plus a conscious lock-in. It produces the same voice.md file in the same shape, so everything downstream keeps working. If a voice.md already exists, or the owner pastes one back at you, you are deepening it, not starting over: Step 3 owns the never-clobber rule, and knowledge/conversation-method.md §6 owns recognising a returning owner and switching to refinement.

The method reference for tone and register is knowledge/communication-voice.md, read it if you haven't.

Step 1: Read the owner's real writing

Voice is captured from real words the owner already wrote, never invented for them. Ask for the richest evidence they can hand you, in this order:

1. A file or folder they point you at, a batch of sent emails, exported posts, a doc of past captions, saved messages, anything they have written. Convert each to clean text first:

   python ~/.claude/bos-run.py tool markitdown_convert "<path-to-file>"

   (The ~/.claude/bos-run.py launcher resolves the install location for you. If it is missing, run python tools/setup.py once from the BOS directory to create it.) This handles PDF, Word, Excel, PowerPoint, images (OCR), HTML, CSV, JSON. If the converter reports it is not installed, relay its one-line install hint and continue with whatever was pasted.

2. Anything they paste, a few emails, posts, or texts dropped straight in.
3. One good prompt when the well is shallow, "Paste me 3-4 emails or posts you wrote and were happy with, and I'll pull out how you sound." A voice built from a handful of genuine samples beats a guessed one.

Read every sample end-to-end. Do not skim. In email threads and message logs, keep the OWNER's lines and drop the other person's, so you capture how they write, not their correspondent.

As you read, gather candidate traits, each with a real example from the samples: typical sentence length and rhythm, vocabulary they reach for, signature moves, greetings and sign-offs, punctuation habits, emoji use, contractions, level of formality, humour, and the registers they never touch. Hold these as observations to confirm, not conclusions to assert.

If there is genuinely almost nothing to read, you can still run Step 2 from how they describe themselves. Say so plainly rather than inventing a voice.

Step 2: The "this, not that" lock-in

This is what makes the voice theirs: a short, targeted session that turns your observations into conscious choices. Draw each contrast from what you actually saw in Step 1.

Run the session under knowledge/conversation-method.md §2-§3: either/or options led by the smart default the samples already support, never an open question. That shape, in this skill's terms: "Your emails open warm and personal, like 'Hope you're having a cracker of a week.' Lock that in as your default open, or do you want something crisper for business notes?"

• Make them choose the boundaries, not just the preferences: which words or registers are off-limits (hype, jargon, corporate filler), and which normal words are absolutely fine so nothing gets over-restricted later.

Contrasts worth locking (pick the ones the samples make live):

• Warm and conversational vs crisp and professional (and whether that shifts by channel)
• Short and punchy vs fuller, more explanatory
• Contractions on vs off; emoji on vs off
• How you open a post: outcome-first vs question/curiosity-first
• Preferred greeting and sign-off
• Signature phrases they own vs phrases they never want put in their mouth
• Jargon tolerance: what's welcome, what's banned

Keep it tight: at most five contrasts in one sitting. That is this skill's declared cap, and the list above is longer than five on purpose, so you are picking the live ones, not walking it. At the fifth you write the file with anything unresolved noted, rather than asking on. Every sample you write during this session follows the content guardrails in the Hard rules below.

Step 3: Write the voice file

Write the company voice to marketing-strategy/<BrandName>/voice.md and the personal voice (if built) to marketing-strategy/<BrandName>/personal-voice.md. Tell the owner exactly where each file is.

Use the same section shape the rest of the system expects, so every content app reads it cleanly:

• Meta-rule (open with it): The owner's voice IS the voice. Everything here is reference for writing in that voice when the owner isn't writing personally. When this doc and the owner disagree, the owner wins, and the doc gets updated to match what they actually shipped.
• Tone, 5 adjectives, each backed by a real quote from the samples.
• Signature moves, the recurring patterns that make them sound like them, with examples.
• Vocabulary to reach for, phrases lifted verbatim from the samples, in a table with a source tag. Available, never mandatory.
• Watch out for, the registers and words the owner chose to avoid in Step 2, framed as defaults not absolute bans.
• Words that are fine, normal words the owner uses, listed so nothing gets over-restricted.
• Mechanical preferences, greeting, sign-off, emoji, punctuation.
• Canonical examples, at least one real sample the owner approved, with a short "why this works" note.
• Change log, dated entries whenever the owner corrects the voice later.

Before you write, run the sharpen pass (conversation-method.md §4) field by field against the evidence rule in the Hard rules below: anything that fails gets cut or labelled, never smoothed over.

Never clobber a hand-tuned file. If voice.md already exists, read it, and show the owner a short before/after of what you are changing before you write. The personal-voice file is framed the same way, in the first person: how the owner writes as themselves.

Leave a pointer in the profile so a later session finds the voice. Once a voice is locked, add one dated line to the ## How the business is running block in ./CLAUDE.md so a returning session (Day 4 of the 5-day challenge reads this block) knows the voice exists and where it lives. This touches only the profile, never the voice files themselves, and follows the same never-clobber discipline:

1. Append, never overwrite. Add the line to the ## How the business is running block (create that heading at the end of the file only if it is missing). Leave every existing line in place. If a pointer line for this voice already exists from a prior run, update it in place rather than adding a duplicate.
2. Use the session's current date. Read today's date at runtime; never hardcode it.
3. Line format: Voice locked (<date>): <short characterisation>, see <voice file path> Example: Voice locked (2026-01-15): direct, warm, short sentences, see marketing-strategy/<BrandName>/voice.md
4. One line per voice. If the personal voice is also locked, add its own pointer line the same way, pointing at personal-voice.md.

Step 4: Prove it, then hand it off

Show the voice working immediately:

• Draft one short sample in each voice you built (for example a post opener and an email line). Let the owner correct it. A correction is not a failure: fold it straight into the file and add a change-log line, so the voice gets truer every time they touch it.
• Then tell them plainly: "That's locked in. From now on, whenever I write a post, a caption, a proposal, or an email, it comes out in this voice. Want me to plan a week of content or draft a post now to see it in action?" Offer the natural next app (write-post-copy, plan-my-content), never push.

Keep your spoken report short: the voice files are the deliverable.

Hard rules

• Ground every trait in real evidence. Do not invent how the owner sounds. If a trait has one supporting sample, say so; if there is no evidence for something, leave it out rather than guessing.
• The owner always wins over the doc, and their correction updates the file (with a change-log entry).
• Content guardrails. Customer-facing copy uses no em dashes, invents no facts, quotes, or numbers, and names no third-party vendor. Write it in the owner's brand voice; the framing and the marketing psychology are the owner's choice. The rules are in knowledge/content-rules.md. The voice files themselves are internal reference.
• Private content stays internal. Real names, private email content, and anything sensitive from the samples belong only in the owner's own files, never carried into public copy.
• Do not clobber an existing voice file without showing the diff first.
• This is the owner's voice, not the customer's. For the words customers use, that is build-customer-voice, a separate input.