✍️ A3: Brand Voice Coach¶

Primary Feature: Instructions & Persona (no knowledge sources) Status: ✅ Built (Corp) Built in: Corp ✅ 2026-04-08 | Lab ⬜

�� Problem It Solves¶

"Our team's communications are inconsistent — every email, announcement, and response sounds different. We need a consistent brand voice across all touchpoints."

This agent handles multiple real-world scenarios:

Customer complaint response (angry, confused, or happy customers)
Corporate announcement rewrites (dry memo → human message)
Internal policy humanising (harsh rules → respectful guidelines)
Sentiment analysis and appropriate response drafting
Multi-audience adaptation (one report → exec / customer / engineering versions)

Why This Agent?¶

This is our Instructions & Persona showcase. It demonstrates that Agent Builder can create a powerful, useful agent using ONLY instructions — zero knowledge sources, zero file uploads, zero web URLs. The entire agent runs on the quality of its instructions.

Key Contrast with A01 and A02¶

Aspect	A01: NZ Policy Advisor	A02: Team Wiki Bot	A03: Brand Voice Coach
Knowledge	🌐 4 web URLs	📂 SharePoint site	❌ None
Power source	Web grounding	Document indexing	Instructions alone
Demo point	External data access	Internal data access	Persona & behaviour control

🏗️ Architecture¶

┌──────────────────────────────────────────────────────────────┐
│                   User Input (any of these)                    │
│                                                                │
│  📧 Customer complaint email (pasted or attached)             │
│  📢 Corporate announcement to rewrite                        │
│  📋 Harsh internal policy to humanise                        │
│  📊 Technical report to adapt for multiple audiences          │
└──────────────────────┬───────────────────────────────────────┘
                       │
                       ▼
┌──────────────────────────────────────────────────────────────┐
│              Brand Voice Coach Agent                           │
│                                                                │
│  📝 Instructions ONLY (no knowledge sources)                  │
│                                                                │
│  ┌──────────────────────────────────────────────────────┐    │
│  │  Brand Voice Definition                               │    │
│  │  • Warm but not casual                                │    │
│  │  • Professional but not stiff                         │    │
│  │  • Clear but not blunt                                │    │
│  │  • Kiwi but not cliché                                │    │
│  ├──────────────────────────────────────────────────────┤    │
│  │  Scenario Handlers                                     │    │
│  │  A: Complaint Response  D: Sentiment Analysis          │    │
│  │  B: Announcement Rewrite E: Multi-Audience             │    │
│  │  C: Policy Humaniser                                   │    │
│  ├──────────────────────────────────────────────────────┤    │
│  │  Boundary Rules                                        │    │
│  │  • NEVER write from scratch                            │    │
│  │  • ONLY rewrite/transform existing text                │    │
│  │  • ALWAYS show before/after comparison                 │    │
│  └──────────────────────────────────────────────────────┘    │
│                                                                │
└──────────────────────┬───────────────────────────────────────┘
                       │
                       ▼
┌──────────────────────────────────────────────────────────────┐
│  📊 Sentiment: 🔴 Angry                                      │
│  📝 Original: [summary of input]                              │
│  ✨ Rewritten in brand voice: [transformed text]              │
│  🔍 What I changed: [3 bullet points]                        │
│  💡 Want me to adjust? [offer to tune]                       │
└──────────────────────────────────────────────────────────────┘

✅ Capabilities Used¶

Capability	Used?	Details
SharePoint Knowledge	❌	Deliberately not used — this agent proves instructions alone are enough
Web Grounding	❌	Not used
File Uploads	❌	Not used (users attach files ad-hoc in chat instead)
Graph Connectors	❌	Not used
Office Artifacts	❌	Not used
Tool Groups	❌	Not used
Copilot Actions	❌	Not used
Instructions	✅	The ONLY feature used — detailed multi-scenario persona with brand voice definition
Starter Prompts	✅	5 working scenarios + 1 boundary test

🔧 Agent Specification¶

Agent Name¶

A03: Brand Voice Coach

Agent Description¶

Rewrites customer emails, announcements, policies, and any text in our warm professional brand voice — detects sentiment and adapts accordingly

🛠️ Quick Build Guide (Copy-Paste Ready)¶

Step	Action
1	Go to microsoft365.com/chat → Agents → Create Agent → Skip to configure
2	Name: `A03: Brand Voice Coach`
3	Description: `Rewrites customer emails, announcements, policies, and any text in our warm professional brand voice — detects sentiment and adapts accordingly`
4	Instructions: Paste the full instruction block from below
5	Knowledge: NONE — leave all knowledge sources OFF (this is the point!)
6	Search all websites: OFF
7	Generate documents, charts, and code: OFF
8	Starter Prompts: Add all prompts from the Starter Prompts table
9	Create and test with T1 (angry customer) first

Demo Files to Have Ready (attach during demo, NOT uploaded as knowledge)¶

File	Location	Scenario
`scenario-01-angry-customer-complaint.txt`	`a03-brand-voice/`	🔴 Furious customer
`scenario-02-dry-corporate-announcement.txt`	`a03-brand-voice/`	📢 Office relocation memo
`scenario-03-confused-customer-inquiry.txt`	`a03-brand-voice/`	🟡 Billing confusion
`scenario-04-happy-customer-praise.txt`	`a03-brand-voice/`	🟢 Team praise
`scenario-05-harsh-internal-policy.txt`	`a03-brand-voice/`	📋 Monitoring policy
`scenario-06-incident-report-multi-audience.txt`	`a03-brand-voice/`	👥 Technical incident

📝 Instructions Used¶

You are the Brand Voice Coach — an expert communication assistant that transforms any text into our company's brand voice. You handle multiple scenarios: customer complaint responses, corporate announcements, policy rewrites, sentiment analysis, and multi-audience adaptation.

## Our Brand Voice: "Professional Warmth"
We sound like a trusted Kiwi advisor who genuinely cares. We are:
- WARM but not casual — "we're here to help" not "yo, hit us up"
- PROFESSIONAL but not stiff — "let's get this sorted" not "please be advised that"
- CLEAR but not blunt — simple explanations without being condescending
- CONFIDENT but not arrogant — we know our stuff but stay humble
- KIWI but not cliché — natural NZ English, use "colour", "organise", "centre"

## Tone Rules
1. Use "we" and "our" — we're a team, not a faceless corporation
2. Use contractions — "we're", "you'll", "here's"
3. Use active voice — "We'll send your report Monday"
4. Start with empathy when addressing problems
5. End with a next step or offer of help
6. Keep sentences short — 15-20 words max
7. NZ English spelling throughout

## Scenario Handling

### Scenario A: Customer Complaint Response
When given an angry or upset customer email:
1. Detect the sentiment and severity (frustrated, angry, furious)
2. Draft a response that: opens with genuine empathy, acknowledges the specific issue, takes responsibility without blame, offers a concrete next step, closes warmly
3. Show a sentiment indicator: 🟢 Happy | 🟡 Neutral | 🟠 Frustrated | 🔴 Angry

### Scenario B: Corporate Announcement Rewrite
When given a dry corporate memo or announcement:
1. Transform it into something people want to read
2. Lead with what matters to the reader (not the company)
3. Use a conversational but professional tone
4. Add a human touch without losing the key information

### Scenario C: Policy Humaniser
When given a harsh or legalistic internal policy:
1. Rewrite to be firm but respectful
2. Explain the "why" behind the policy
3. Focus on benefits, not threats
4. Replace "failure to comply" language with collaborative framing

### Scenario D: Sentiment Analysis and Response
When given any customer communication:
1. Analyse the sentiment (with indicator emoji)
2. Identify the key concerns or questions
3. Draft an appropriate on-brand response
4. Suggest follow-up actions

### Scenario E: Multi-Audience Adaptation
When given a technical or internal document and asked to adapt:
1. Rewrite for the requested audience (executive, customer, team, public)
2. Adjust detail level, tone, and jargon accordingly
3. Maintain accuracy while changing the style

## Response Format
Always structure your response as:

📊 Sentiment: [emoji indicator] [sentiment description]
📝 Original: [summarise or quote the key text]
✨ Rewritten in our brand voice:
[your rewrite]
🔍 What I changed and why:
- [change 1]
- [change 2]
- [change 3]
💡 Want me to adjust? I can make it more formal, more casual, shorter, longer, or adapt it for a different audience.

## What You Do NOT Do
- NEVER write content from scratch. If a user asks you to create original content without providing existing text to rewrite, respond: "I'm your Brand Voice Coach — I transform existing text into our brand voice. Please give me a draft, email, announcement, or document to work with, and I'll make it shine! ✨"
- NEVER change facts or add information not in the original
- NEVER handle topics outside communication and writing. If asked, respond: "I specialise in brand voice and communication — that question is outside my area. Try asking Copilot directly!"

🚨 Gotchas & Design Decisions¶

📝 Gotcha 1: This Agent Has ZERO Knowledge Sources — And That's the Point¶

Most people assume every agent needs data to be useful. A03 proves that instructions alone can create a genuinely valuable agent. The entire brand voice definition, scenario handling, and response formatting lives in the instructions.

🎤 Demo Talking Point

"This agent has zero knowledge sources — no SharePoint, no files, no websites. Everything it does comes from the instructions I wrote. That's how powerful good instructions are. You can build a useful agent in 5 minutes with nothing but a well-written prompt."

🔴 Gotcha 2: Soft Boundaries Get Ignored — Use the CAPS Technique¶

What happened: Our first version had "Don't write content from scratch without a draft" as a boundary rule. When tested, the agent ignored it and happily wrote a marketing email from scratch.

Why: The LLM's instinct to be helpful overrides gentle "don't" instructions. It sees a user asking for help and wants to deliver.

The fix: Three-part boundary enforcement:

Technique	Example	Why It Works
1. ALL CAPS keywords	`NEVER write from scratch` instead of `Don't write from scratch`	LLM treats CAPS as higher priority signals
2. Exact fallback text	`respond: "I'm your Brand Voice Coach — I transform existing text..."`	Removes temptation to improvise a "helpful" answer
3. Positive redirect	`"Please give me a draft to work with, and I'll make it shine!"`	Satisfies the LLM's helpfulness instinct while enforcing the boundary

⚠️ Lesson Learned

Always test your boundary prompts BEFORE demoing. We caught this during our build, not during a customer demo. Every "Topics You Do NOT Cover" rule should have a matching test prompt.

This technique is now documented in the main Agent Builder overview as a best practice for all agents.

📎 Gotcha 3: File Attachments Work in Agent Chat¶

Users can attach files directly in the Copilot Chat when talking to an agent. You don't need to configure File Upload as a knowledge source — ad-hoc attachments are a chat feature.

This means for demos: - Prepare scenario files in advance (we have 6 ready) - Click the attach/paperclip icon in the agent chat - Upload the scenario file - The agent reads it and applies the brand voice

File Attachments vs Knowledge Sources

Feature	File Attachments (in chat)	Knowledge Sources (configured)
How added	User attaches during conversation	Creator configures during build
Persistence	Just for that conversation	Always available to the agent
Limit	Ad-hoc, per conversation	Up to 20 files / 20 SP sites
Use case	One-off document processing	Permanent reference data

🎨 Gotcha 4: Brand Voice Definition Takes Up Instruction Space¶

Our brand voice definition (the "Professional Warmth" section with DO/DON'T table, tone rules, scenario handlers) uses roughly 3,500 of the 8,000 character limit. That leaves ~4,500 characters for rules, scenarios, and boundaries.

Trade-off: More detailed brand voice = better consistency, but less room for scenario handling and rules.

Our decision: We balanced detail (enough for the LLM to understand the voice) with breadth (5 scenario handlers + boundary rules). If you need more scenarios, consider splitting into multiple specialised agents.

💬 Starter Prompts¶

#	Title	Prompt	Scenario
T1	🔴 Angry Customer	I'm going to attach an angry customer complaint. Please draft a response in our brand voice.	A: Complaint Response
T2	📢 Announcement	Rewrite this corporate announcement to be more human: "Please be advised that effective immediately, all staff are required to complete mandatory compliance training by April 30. Failure to complete training will result in system access revocation."	B: Announcement Rewrite
T3	📋 Policy	This policy sounds too harsh. Make it warmer: "Employees found using personal devices for work purposes without authorisation will be subject to disciplinary proceedings."	C: Policy Humaniser
T4	🎭 Sentiment	Analyse the sentiment of this message and draft a response: "I've been waiting 2 weeks for my refund. This is getting ridiculous."	D: Sentiment Analysis
T5	👥 Multi-Audience	I'm going to share a technical incident report. Please rewrite it for three audiences: the executive team, affected customers, and our internal engineering team.	E: Multi-Audience
B1	🚫 From Scratch	Write me a marketing email from scratch	Boundary — should decline

📁 Demo Scenario Files¶

Pre-built files at C:\ssClawy\agent-builder-demo-content\a03-brand-voice\:

#	File	Scenario	Sentiment	Demo Impact
1	`scenario-01-angry-customer-complaint.txt`	3-week ignored customer threatening to cancel	🔴 Furious	⭐⭐⭐ Highest — great opener
2	`scenario-02-dry-corporate-announcement.txt`	Office relocation memo with "failure to comply" language	😐 Dry/corporate	⭐⭐ Good contrast
3	`scenario-03-confused-customer-inquiry.txt`	Billing confusion + can't find features	🟡 Confused	⭐⭐ Shows gentle handling
4	`scenario-04-happy-customer-praise.txt`	Customer praising a specific team member	🟢 Happy	⭐ Shows positive response handling
5	`scenario-05-harsh-internal-policy.txt`	Employee monitoring/tracking policy	😠 Harsh/threatening	⭐⭐⭐ Great for policy humanising
6	`scenario-06-incident-report-multi-audience.txt`	P2 portal outage — technical details	📊 Technical/neutral	⭐⭐⭐ Shows multi-audience versatility

💡 Recommended Demo Order

For maximum impact: Scenario 1 (angry customer) → Scenario 5 (harsh policy) → Scenario 6 (multi-audience). Start with the biggest wow factor, end with versatility.

🧪 Test Scenarios¶

✅ In-Scope Tests¶

#	Test	How to Test	Expected Behaviour	Actual Result	Pass?
T1	Angry customer response	Attach scenario-01 file	Detect 🔴 Angry sentiment, draft empathetic response, show before/after	✅ Worked	✅
T2	Announcement rewrite	Use starter prompt T2	Transform corporate → human, keep facts, change tone
T3	Policy humaniser	Use starter prompt T3	Soften language, add "why", remove threats
T4	Sentiment analysis	Use starter prompt T4	Detect 🟠 Frustrated, draft response, suggest next steps
T5	Multi-audience	Attach scenario-06 file	Three distinct versions: exec (brief), customer (empathetic), engineering (detailed)

🚫 Boundary Tests¶

#	Test Prompt	Expected Behaviour	Actual Result	Pass?
B1	"Write me a marketing email from scratch"	DECLINE with exact fallback text + redirect	❌ First attempt: agent wrote the email (soft boundary failed) → ✅ Fixed with CAPS technique	✅ (after fix)
B2	"What is the weather in Auckland?"	DECLINE — outside communication/writing scope

👍 Pros & Cons¶

Pros	Cons
✅ Zero setup — no knowledge sources needed at all	❌ Brand voice quality depends entirely on instruction quality
✅ Built in under 5 minutes	❌ 8,000 char limit means trade-off between voice detail and scenario count
✅ File attachments in chat enable rich demo scenarios	❌ Soft boundaries fail — need CAPS technique (learned the hard way)
✅ Multiple scenarios in one agent (complaint, announcement, policy, etc.)	❌ No memory between conversations — can't "learn" preferences over time
✅ Instantly relatable to every business	❌ Agent may occasionally deviate from voice on edge cases
✅ Perfect showcase for "instructions are the #1 thing to get right"	❌ NZ English spelling rule sometimes ignored by the LLM

🔧 Workarounds Applied¶

Issue	Workaround	Rationale
Boundary rule ignored (B1 failed)	CAPS keywords + exact fallback text + positive redirect	Three-part technique: `NEVER` + scripted response + redirect to what user should do instead
8,000 char limit tight with 5 scenarios	Balanced brand voice detail vs scenario count	Enough voice definition for consistency, enough scenarios for versatility
Users need to provide draft text	Added "attach file" workflow to demo	More professional than copy-pasting text into chat
NZ English spelling sometimes ignored	Added explicit spelling rule in instructions	"Use NZ English — colour, organise, centre (not US spelling)"

🔮 Future Enhancements¶

Enhancement	Feature Needed	Impact
Add company style guide as knowledge	SharePoint or File Upload	Agent references actual brand guidelines, not just instructions
Generate rewritten docs as Word files	Office Artifacts	Output polished Word documents, not just chat text
Auto-process incoming emails	Copilot Actions + Outlook Tool Group	Automatically draft responses to customer emails
Add writing examples as reference	File Upload knowledge	Show agent examples of "good" brand voice for better consistency
Multi-language support	Instructions update	Add Maori greetings/phrases for NZ cultural touch

📸 Screenshots¶

(Add screenshots from Corp and Lab builds)

Corp Build¶

(pending)

Lab Build¶

(pending)

🎤 Demo Talking Points¶

"Zero knowledge sources" — This agent has no SharePoint, no files, no websites. Everything comes from the instructions. That is how powerful good instructions are.
"Built in under 5 minutes" — Show the timer. Name, description, instructions, prompts. Done.
"Attach and transform" — Drop a customer complaint email into the chat. Watch the agent detect the sentiment and draft a warm, professional response.
"Boundary enforcement" — Show B1 (from-scratch request). The agent politely declines and redirects. Your instructions control what it does AND what it refuses to do.
"The CAPS technique" — Tell the customer: "We learned that soft rules get ignored. Use NEVER and ALWAYS in caps for critical boundaries. It works."
"Five scenarios, one agent" — Complaints, announcements, policies, sentiment, multi-audience. All from the same set of instructions.
"Imagine combining this with SharePoint" — This agent + your actual brand style guide on SharePoint = bulletproof brand consistency.

← Back to Agent Builder Overview