Draft Setup

Rich onboarding to calibrate the draft to your authentic voice. Scans your messaging history, collects samples, asks questions, then saves a voice profile and style examples to kbx for future sessions.

Voice

Conversational and collaborative. You're learning about the user's voice, not lecturing them. One question at a time.

Process

1. Pre-Check

Check for an existing profile:

Check if memory/draft/voice-profile.md exists (Read tool)

Condition	Action
Profile exists	Show current profile summary, offer: update, start fresh, or cancel
No profile	Proceed with setup

2. Welcome

Let's calibrate your voice. This takes about 5-10 minutes — I'll scan your recent messages, ask some questions, and build a voice profile.

At the end, I'll save everything so future drafting sessions sound like you.

3. Phase 1 — Deep Scan (background)

Spawn a background worker to scan the user's messaging history:

Slack scan (if Slack MCP available):

• slack_search_public for recent messages from the user across channels
• Aim for 20-30 representative messages across different contexts (announcements, threads, praise, decisions)
• Categorise each message by apparent style

Gmail scan (if Gmail MCP available):

• gmail_search_messages for recent sent emails
• Focus on substantive messages (skip auto-replies, one-liners)
• Collect 5-10 representative emails

While the scan runs, proceed to Phase 2.

4. Phase 2 — User Samples

Present scan findings when available:

I found [N] messages across [M] contexts. Here are the ones that seem most distinctive:

[Show 3-5 most characteristic messages with detected style labels]

Then ask:

Do you have 2-3 additional messages that feel most "you"? Paste them here or share Slack links.

These are the messages where someone who knows you would say "that's definitely [Name]."

Accept:

• Pasted text: Use directly
• Slack links: Read via slack_read_thread or slack_read_channel MCP
• "Skip": Proceed with scan data only

5. Phase 3 — Q&A (one question at a time)

Ask these in sequence. Use AskUserQuestion with multiple-choice where possible. Skip questions that are already clearly answered by the scan data.

Q1: Role and audience

What's your role, and who do you primarily communicate with? (This helps me calibrate formality and authority level.)

Q2: Tone (3 words)

How would you describe your communication style in 3 words?

Q3: Language

Which English spelling convention do you use?

• British English (colour, organise, behaviour)
• American English (color, organize, behavior)
• Other / mixed

Q4: Greetings

How do you typically open group messages? (e.g., "Hi folks", "Hey team", "Good morning everyone")

Q5: Sign-offs

How do you close messages? (e.g., just your name, "Thanks!", "Have a great weekend!", no sign-off)

Q6: Anti-patterns

Any words or phrases you hate or want to avoid? Corporate jargon, particular expressions, anything that makes you cringe when you see it in a message?

Q7: Emoji

How do you use emoji?

• Sparingly and naturally (1-2 per message max)
• Frequently (multiple per message, emoji reactions)
• Rarely / never

Q8: Formatting

Do you prefer bullets, prose, or mixed? Do you use bold for emphasis?

Q9: Overused phrases

Any patterns you tend to overuse that we should vary? (Most people have 2-3.)

Q10: Sensitive topics

Any topics that need special care? (Sensitive projects, specific people, political subjects within the org.)

6. Phase 4 — Analysis

Synthesise all data (scan + samples + Q&A) into a voice fingerprint:

1. Analyse patterns across all collected messages:

   • Sentence length and rhythm
   • Greeting/sign-off patterns
   • Emoji frequency and types
   • Formatting preferences (bullets vs prose, bolding)
   • Vocabulary patterns (favourite words, phrases)
   • Communication philosophy (BLUF? storytelling? question-led?)

2. Detect communication styles from the corpus:

   • Map messages to the 8 default categories (Strategic Announcement, Action-Oriented, etc.)
   • Flag categories with no examples ("You don't seem to use Knowledge Share style often")
   • Identify any styles not covered by the defaults ("You have a distinct 'weekly reflection' style")

3. Present the profile:

Here's what I've learned about your voice:

Identity: [role, tone description] Language: [spelling, greetings, sign-offs] Writing cadence: [sentence rhythm, paragraph style] Formatting: [emoji, bold, bullets preferences] Anti-patterns: [words to avoid] Sensitive topics: [areas needing care]

Detected styles (with example counts):

• Strategic Announcement: [N] examples
• Action-Oriented: [N] examples
• [etc.]
• [Custom style if detected]: [N] examples

Does this capture your voice? Anything to adjust?

7. Phase 5 — Review & Save

Accept adjustments, then save:

1. Voice profile — write to memory/draft/voice-profile.md:

   • Use the format defined in skills/voice-identity/SKILL.md
   • Include all sections: Identity, Language, Anti-Patterns, Writing Cadence, Formatting, Sensitive Topics

2. Style files — write to memory/draft/styles/<style-name>.md for each detected style:

   • YAML frontmatter: title, type (draft-style), style slug, tags
   • Voice Notes section: style-specific observations
   • Examples section: the user's real messages for this style with annotations

3. Confirm:

Voice profile saved with [N] style files.

You can update anytime by running /draft:setup again.

Ready to draft something? Run /draft:boot or just tell me what you need.

Graceful Degradation

Missing	Fallback
Slack MCP unavailable	Skip deep scan. Ask user to paste 5-10 sample messages instead.
Gmail MCP unavailable	Skip email scan. Slack + pasted samples are enough.
Both MCPs unavailable	Full manual mode — rely on pasted samples + Q&A only.
User skips questions	Work with whatever they provide — any data is better than none.
kbx/file write fails	Present the profile as formatted text for the user to save manually.