Building a Custom GPT for Your Support Team
Custom GPTs (and Claude Projects) are one of the highest-ROI AI moves a support team can make. Instead of every agent writing long prompts from scratch, you create one specialized assistant pre-loaded with your brand voice, policies, and knowledge base. Every agent gets a personal expert trained on your company for free. This lesson walks through building one end-to-end.
What You'll Learn
- The difference between Custom GPTs, Claude Projects, and Gems
- Step-by-step: building a "Support Agent Copilot" in 30 minutes
- What to put in the instructions, knowledge files, and capabilities
- Sharing it with your team and keeping it current
What a Custom GPT Actually Is
A Custom GPT is a ChatGPT chatbot with:
- Custom instructions -- a system prompt telling it who it is and how to behave
- Knowledge files -- PDFs, Word docs, text files it can reference (up to 20 files per GPT)
- Capabilities -- web browsing, image generation, code interpretation (on/off per GPT)
- Actions -- optional API connections to your systems
It runs in ChatGPT's interface. Anyone with the link (or the whole public Custom GPT store) can use it.
Claude Projects is the equivalent in Claude -- a project has instructions and a knowledge base and lets multiple people chat with that shared context. Google's Gems is the Gemini equivalent.
All three serve the same use case, and you'd pick based on which AI tool your team already uses.
Why a Custom Support GPT Is Worth Building
Once built, every agent on your team:
- Has instant access to an AI that knows your brand voice
- Doesn't need to paste context into prompts every time
- Gets consistent reply quality across the team
- Onboards in days instead of weeks (new agents use the GPT as a training wheel)
Teams building these report saving 3-5 hours per agent per week within the first month. And it's free if you have a ChatGPT Plus or Team subscription.
Step-by-Step: Building Your Support GPT
Step 1: Decide What It Does
You want a narrow, well-defined purpose. Good purposes:
- "Draft reply to incoming ticket following our brand voice"
- "Triage and summarize incoming tickets"
- "Write/update KB articles"
Bad purpose: "Do everything support-related." Narrow GPTs outperform broad ones.
For this lesson, we'll build a Support Reply Copilot: paste a ticket, get a draft reply in our brand voice, using our policies, in under 130 words.
Step 2: Create It
In ChatGPT (Plus/Team/Enterprise):
- Click your profile -> "My GPTs" -> "Create a GPT"
- Use the "Configure" tab (skip the conversational setup for faster control)
- Give it a name: "[Company] Support Copilot"
- Give it a description: "Drafts customer support replies in our brand voice. Use when you need a polished first draft to edit and send."
Step 3: Write the Instructions
This is the core of your GPT. The template below is a great starting point:
You are the Support Reply Copilot for [Company], a [brief company description].
You help support agents draft polished, on-brand replies to customer tickets.
WHEN AGENT PASTES A TICKET:
- Ask clarifying questions if essential context is missing (customer name, plan, the issue)
- Otherwise, produce a draft reply immediately
OUTPUT FORMAT:
- Subject line (6-10 words, no "Re:")
- Email body under 130 words
- Include {{first_name}} if agent didn't provide the customer's name
BRAND VOICE:
- Warm, confident, uses contractions
- Starts with an acknowledgement of the customer's feeling
- Never uses: "unfortunately," "per our policy," "we value your feedback," "as previously stated"
- Uses "I'll fix this" not "I'll try to fix this"
- Signs off warmly
ACCURACY RULES:
- Never invent policies, prices, features, or refund amounts
- If the agent hasn't supplied a specific number, use a placeholder like [refund amount]
- If the question needs information not in your knowledge files, say "I don't have that info -- the agent should verify with engineering"
POLICIES:
[paste key policies: returns, refunds, subscriptions, etc.]
NEVER:
- Draft replies for legal matters, data breaches, bereavement, or threats
- Invent specific dates for bug fixes or product updates
- Contradict the knowledge files
This instruction block is the single most valuable thing you'll write. Polish it over a few iterations with your team.
Step 4: Upload Knowledge Files
Upload:
- Your brand voice guide (1 PDF)
- Your support policies (1 PDF)
- Your top 50 knowledge base articles (combined into 1-2 PDFs or Markdown files)
- Your product overview (1 PDF)
- Your current pricing (1 PDF)
Keep it under 20 files, ideally 5-10 high-quality ones. Large, well-organized files beat many tiny ones.
Step 5: Set Capabilities
Recommended settings:
- Web browsing: Off (prevents hallucinations from random sites)
- DALL-E image generation: Off
- Code interpreter: Off unless you use it for ticket analysis
You want the GPT grounded in your files, not the internet.
Step 6: Test With Real Tickets
Test with 10-15 real tickets from your queue (anonymize any sensitive data). Score each output:
- Tone alignment?
- Accuracy?
- Correct policy citations?
- Correct length?
Iterate the instructions based on failures. Most teams iterate 3-5 times before the GPT is reliably good.
Step 7: Share With the Team
Set sharing to:
- "Anyone with the link" if you trust your team to keep the link private
- "Only me" if you want to gate access
- Or if on ChatGPT Team, share to your Team workspace
Walk through it with the team in a 15-minute demo. Show them what it does well and its limitations.
Claude Projects: The Equivalent
If your team uses Claude (claude.ai), create a Project:
- Claude -> Projects -> New Project
- Name it, describe it
- Add Project Knowledge (files + text snippets)
- Set Custom Instructions (similar structure to above)
- Share with your team via Claude's team features
Claude Projects often handle nuanced tone better out of the box. The file handling is similar.
What to Include in Your Knowledge Base
Good knowledge additions:
- Policy documents (returns, privacy, pricing)
- Brand voice / tone guide
- Top 50 KB articles
- Common scenario playbooks (chargebacks, cancellations, escalations)
- Past "exemplar" replies that scored high on CSAT
Avoid:
- Personal customer data (PII)
- Internal-only pricing or deals
- Confidential product roadmap
- Anything your legal team wouldn't want leaked
Remember: anyone with access to the GPT can potentially extract its knowledge via clever prompts.
Maintaining the GPT Over Time
Set a quarterly calendar:
- Monthly: Re-upload the updated KB file
- Quarterly: Review the instructions block, update with new policies or brand voice evolutions
- Whenever a policy changes: Update immediately
An outdated GPT is worse than no GPT because agents trust it and get wrong answers.
Other Useful Support GPTs to Build
Once you've built one, build more narrow ones:
- Ticket Triage GPT: Classifies and prioritizes incoming tickets
- KB Article Writer: Takes a resolved ticket, outputs a KB article
- Macro Generator: Takes a scenario, outputs a ready-to-use macro
- CSAT Analyzer: Takes a batch of CSAT comments, returns themes and quotes
- Translation Assistant: Your brand voice, in multiple languages
Narrow GPTs with focused instructions beat one giant GPT that tries to do everything.
Key Takeaways
- Custom GPTs (ChatGPT), Projects (Claude), and Gems (Gemini) all serve the same use: a shared, pre-configured AI assistant
- Start with one narrow GPT -- a Support Reply Copilot is the highest-ROI first build
- The instructions block is the heart of the GPT; iterate it 3-5 times before launching
- Turn off web browsing; ground answers in your uploaded knowledge files only
- Keep the GPT current -- re-upload the KB monthly, review instructions quarterly