Build a Custom GPT for Your Marketing Team in Under 30 Minutes

First published: 20 June 2026 · Last updated: 20 June 2026

1

Define one specific job

Brand Voice Editor, Email Subject Line Generator, FAQ Drafter. One job per GPT. Generic "marketing assistant" GPTs underperform specialised ones every time.

2

Write tight Instructions

200 to 500 words covering tone, format, must-do, must-not-do, and the exact output structure. Be explicit. The GPT will not infer.

3

Upload 3 to 5 knowledge files

Brand voice doc, top-performing past examples, your style guide, customer language samples. Quality over quantity.

4

Test with 10 to 15 real prompts

Use prompts your team will actually send. Iterate Instructions until 80% of outputs are usable with light editing.

5

Share via link with your team

One-click access. Team members do not need a Plus account if you are on Team plan. Track usage and refine quarterly.

Most SG marketing teams I talk to are still using ChatGPT the same way they used Google in 2010: open a fresh tab, type a request, copy what comes back, hope it sounds on-brand. That works for one-off tasks. It does not scale across a team of three to ten marketers all producing copy that needs to sound like the same brand. A Custom GPT solves this. You configure the brand voice, the format, the do-not-do list and the reference materials once, then everyone on the team uses the same tuned version through one link. This is a hands-on tutorial. By the end, you will have a working "Brand Voice Custom GPT" that takes a draft from any team member and rewrites it in your house voice, with consistency across writers. Total time: about 30 minutes if you have your brand voice document ready, 45 if you need to compile one along the way. We covered the broader landscape of AI tools in our best AI marketing tools in Singapore guide; this piece is the build-it-yourself how-to for one of them.

Before You Start: What You Need

Three prerequisites. Get these ready before you open the GPT builder or you will hit a wall halfway through. **ChatGPT Plus, Team or Enterprise account.** Custom GPT building is gated behind paid plans. Plus is USD 20 per month per user. Team is USD 25 per user per month with shared workspaces and is the right choice for any SG marketing team of 2 or more (you get shared GPTs and admin controls). Enterprise makes sense for larger teams with compliance requirements. **A brand voice document or the willingness to write one.** This is the single most important input. If your brand has never written down its voice rules, take 60 minutes before this exercise and answer: what 5 adjectives describe how we sound? What 3 things do we never say? What is the reading age we write for? Who is our reader? What is one paragraph of copy that perfectly represents our voice? You do not need a 50-page brand bible. You need 1 to 3 pages of clear specifics. **3 to 5 examples of your best past content.** Top-performing emails, the homepage hero copy, a blog intro you are proud of, a case study quote, an ad headline that converted. The GPT learns voice as much from examples as from rules. Examples are non-negotiable.

Step 1: Open the GPT Builder

Inside ChatGPT, click your name (bottom left), select "My GPTs", then "Create a GPT". You will see two tabs at the top: Create and Configure. Create is a chat-based wizard where ChatGPT helps you build the GPT through conversation. Configure is the manual editor where you set everything explicitly. Recommended workflow: start in Create for 5 minutes to get a rough draft, then switch to Configure to do the real work. The Create wizard is friendly but produces generic GPTs; you will want to overwrite most of what it generates. In Create, paste this opening prompt: "I want to build a Custom GPT called [Brand] Voice Editor. It takes any draft text and rewrites it in our brand voice. Our voice is [adjective 1], [adjective 2], [adjective 3]. We never use [thing 1] or [thing 2]. Our typical reader is [reader description]. Help me draft the Instructions for this GPT." ChatGPT will produce a first draft. Copy it. Now switch to Configure. We are going to overwrite most of it.

Step 2: Write the Instructions Field (the Most Important 30 Minutes)

The Instructions field is where you tell the GPT exactly how to behave. This is the single highest-leverage thing you do. Get this right and the GPT works. Get this wrong and you will be frustrated. Structure your Instructions in this order. Use clear headings. The GPT reads these every conversation: **1. Role and purpose.** "You are the [Brand] Voice Editor. Your job is to take any draft text the user provides and rewrite it so it perfectly matches our brand voice. You do not generate net-new copy from scratch unless explicitly asked." **2. Voice rules.** Five adjectives. Three things we never say. The reading age. Sentence length preferences. Specific words or phrases we love and ones we hate. Be brutal. "We never use 'leverage' as a verb. We never use 'in today's fast-paced world'. We never use em dashes. We use the Oxford comma." **3. Format rules.** When given a draft, what does the GPT return? "Return the rewritten version first, with no preamble. Then under a horizontal rule, list the 3 main changes you made and why." **4. Must-do list.** "Always preserve the user's original meaning. Always keep numbers, names and links exactly as the user provided them. Always write in Singapore English (colour, organisation, behaviour). Always use SGD when discussing pricing." **5. Must-not-do list.** "Never invent statistics. Never add bullet points unless the original had them. Never use the word 'just' as a softener. Never use exclamation marks except in approved scenarios (greetings, sales urgency)." **6. Worked example.** Paste one before-and-after rewrite so the GPT has a concrete pattern to copy. This single example is worth more than 200 words of abstract instruction. **7. Edge cases.** "If the user provides only a topic and asks for original copy, ask them what format they want (email, ad headline, blog intro) before generating. If the user provides text in another language, translate it to English first, then rewrite in our voice." Aim for 300 to 500 words total in Instructions. Shorter is better than longer. Specific is better than vague. The GPT will not guess what you mean; if a rule is not explicit, it will not be followed consistently.

Step 3: Upload Knowledge Files

In the Configure tab, scroll to "Knowledge". Upload 3 to 5 files. Quality matters far more than quantity here. Recommended uploads for a Brand Voice Editor GPT: **1. Your brand voice document** (PDF or markdown). The full version with examples. **2. A "best of" anthology** (PDF or text). Compile 5 to 10 pieces of your best past copy: top emails, homepage hero, a strong blog intro, a converting ad. Label each one with its context. **3. Your style guide** (PDF). Spelling preferences (Singapore English), punctuation rules, capitalisation conventions, brand name usage, product name capitalisation. **4. A "what not to do" file** (markdown). Real examples of off-brand drafts your team has produced, with annotations explaining why they are wrong. The negative examples are as instructive as the positive ones. **5. (Optional) Customer language samples** (text). Verbatim quotes from customer interviews, support tickets, reviews. Helps the GPT match the language your audience uses. Skip generic marketing books, agency pitches, and competitor content. The GPT will indiscriminately pull from anything you upload, so anything you do not want it to imitate should not be in the knowledge folder.

Upload (do)	Do not upload
Your brand voice doc with examples	Generic marketing or copywriting books
5-10 of your best past copy pieces	Competitor websites or ads
Style guide (Singapore English, brand naming)	Industry trend reports
"What not to do" annotated examples	HR policies or unrelated company docs
Customer interview quotes / reviews	Outdated brand assets from 2 years ago
Approved product / service descriptions	Anything you would not want it to imitate

Step 4: Configure the Capabilities Toggles

ChatGPT lets you toggle which capabilities the GPT has access to. For a Brand Voice Editor GPT:

• Web Browsing: Off. You do not want it pulling fresh content from the web mid-rewrite.
• DALL-E Image Generation: Off (unless this GPT also produces images, which is rare for a voice editor).
• Code Interpreter: Off.
• Actions: Off (unless you are building API integrations, which is a separate advanced topic).

Fewer capabilities equal more predictable behaviour. Only enable what the GPT actually needs.

Step 5: Add Conversation Starters

In Configure, add 4 conversation starters. These appear as clickable buttons when a team member opens the GPT. They guide the team to the most common use cases. For a Brand Voice Editor:

• "Rewrite this email in our voice: [paste email]"
• "Rewrite this ad headline in our voice: [paste headline]"
• "Rewrite this blog intro in our voice: [paste intro]"
• "Audit this draft against our voice rules and tell me what is off-brand"

Conversation starters dramatically increase usage. Team members are 3 to 5x more likely to use the GPT when there are clear example prompts on the opening screen.

Step 6: Test With 15 Real Prompts

Now the actual work begins. Open the preview pane on the right side of the Configure screen. Run 15 prompts that represent real things your team will send. Use actual drafts from recent work, not invented examples.

For each output, score it 1 to 5:

• 5: Publishable as is.
• 4: Usable with under 30 seconds of editing.
• 3: Usable with 1 to 3 minutes of editing.
• 2: Significant rework needed.
• 1: Worse than starting from scratch.

Your goal: 80% of outputs scoring 4 or higher. If you are not there, the Instructions field needs work. The most common fix is being more specific about format and structure (a vague "match our voice" produces less consistent output than "use sentences of 8 to 18 words, mostly in active voice, with one rhetorical question per paragraph").

Iterate. Tighten Instructions. Add a clarifying example. Add an explicit don't. Re-test. Two to three iteration cycles is normal. The work pays off because every team member who uses this GPT thereafter inherits your tuning.

Step 7: Share With Your Team

Once tests pass, click "Share" in the top right. Choose "Anyone with the link" if you are on Plus, or share via your Team workspace if you are on Team. Send the link to your marketing team. Pin it in your team docs or Notion.

Set a quarterly review: every 90 days, ask the team for the 5 prompts that produced bad output. Add fixes to Instructions. Update knowledge files with new best examples from the quarter. Custom GPTs are not "build once and forget"; they compound in usefulness when you tune them based on real team feedback.

What to Build Next: 5 More Custom GPTs Worth the 30 Minutes

Once you have one Custom GPT working, the leverage compounds. Other high-value GPTs for SG marketing teams:

1. Email Subject Line Generator. Knowledge: your top 50 historical email subjects with open rates. Instructions: produce 10 subject line options for a given email, each in a different angle (curiosity, benefit, urgency, contrarian, social proof, question, list, news, personalisation, statement). Singapore SME marketers especially benefit from this because subject line testing is where most SG email programs underperform.

2. Ad Copy Variant Generator. Knowledge: your top 20 historical ads. Instructions: produce 5 ad headline and body copy variants for a given product or offer, optimised for the specified channel (Meta, Google, LinkedIn).

3. SG Customer Persona Interviewer. Knowledge: customer interview transcripts. Instructions: ask 10 to 15 sharp questions about a draft offer, simulating how a specific persona would react. Useful for stress-testing landing pages before launch.

4. SEO Brief Builder. Knowledge: your top-ranking blog briefs and structure docs. Instructions: take a target keyword and produce a complete brief with target word count, suggested H2s, related questions to answer, and 5 internal link recommendations.

5. Singlish-to-English Editor. Knowledge: your style guide on tone for international audiences. Instructions: take Singlish-flavoured copy and rewrite it in standard English while preserving the energy and personality. Useful for SG brands expanding regionally.

For deeper prompt patterns to use across all of these, our AI prompts for marketing guide covers reusable prompt frameworks. For where Custom GPTs sit inside a broader workflow, see our piece on the AI marketing workflow.

1

Email Subject Line Generator

Trained on your top 50 historical subjects with open rates. Outputs 10 variants per request across distinct angles.

2

Ad Copy Variant Generator

Trained on top 20 historical ads. 5 channel-specific variants per request. Faster than briefing a copywriter for low-stakes tests.

3

SG Customer Persona Interviewer

Stress-tests landing page drafts by simulating how a named persona would react. Catches copy weaknesses before launch.

4

SEO Brief Builder

Takes a keyword, returns target length, H2 suggestions, related questions, internal link recommendations. 10 minutes saved per brief.

5

Singlish-to-English Editor

For SG brands going regional. Rewrites Singlish-flavoured copy in standard English without flattening the energy.

Common Mistakes Building Custom GPTs

Five we see SG teams make repeatedly.

Mistake 1: Trying to build a generic "marketing assistant" GPT. One GPT that does everything. The output quality is mediocre across all tasks. Specialised single-purpose GPTs outperform generalist ones every time. Build five specific GPTs, not one general one.

Mistake 2: Vague Instructions. "Write in our brand voice." The GPT has no idea what that means. Be explicit: voice adjectives, sentence length, what to never say, format rules, worked example. Specificity is the difference between a GPT that works and one that frustrates.

Mistake 3: Dumping 50 files into Knowledge. More files do not equal better output. They equal more noise and slower retrieval. Stick to 3 to 5 high-quality files. Curate ruthlessly.

Mistake 4: Skipping the test phase. Building a GPT and immediately sharing it with the team without running 15 test prompts. The team will hit failures, conclude the GPT is useless, and stop using it. Test before you ship.

Mistake 5: Never updating the GPT after launch. Custom GPTs need quarterly tuning. New brand voice rules, new top-performing copy, new common failure modes. A GPT that is not updated for a year drifts away from current brand reality and becomes an obstacle.

Frequently Asked Questions

Do team members need their own ChatGPT Plus account to use a Custom GPT I built?

If you are on ChatGPT Plus and share via "Anyone with the link", they need a Plus account to use it. If you are on ChatGPT Team or Enterprise, you can share Custom GPTs internally and team members can use them through the team workspace without their own Plus account. For any SG marketing team of 2 or more, the Team plan (USD 25 per user per month) is the right call.

How is a Custom GPT different from just using a long system prompt every time?

The Custom GPT bakes the system prompt, knowledge files and conversation starters into a single shareable interface. Every team member gets the same configuration, no copy-paste of prompts, no version drift. You also get built-in knowledge file retrieval that you cannot replicate with a system prompt alone. Functionally similar but operationally far more scalable.

Can my Custom GPT see or learn from team conversations over time?

No. Each conversation is independent. The GPT does not learn from team usage. The only way to update the GPT is for you (the builder) to edit Instructions or update knowledge files manually. This is good for predictability but means you have to commit to quarterly tuning to keep the GPT improving.

Are my knowledge files visible to other ChatGPT users?

Files you upload to your Custom GPT's knowledge are visible only to people who use your GPT, not to the wider ChatGPT user base. However, content from those files can appear in GPT responses, so do not upload anything you would not want quoted back. Avoid uploading confidential client data, NDAs, or PII.

How does this compare to building on Claude Projects or Gemini Gems?

Functionally similar. Claude Projects and Gemini Gems both offer comparable "configure once, share with team" patterns. Choose based on which underlying model your team prefers and which integrations you need. We covered the broader trade-offs in our ChatGPT vs Claude vs Gemini for marketing comparison. Most SG teams default to ChatGPT for the user experience and the Custom GPT marketplace, but Claude has a usability edge for long-form copywriting work.

What if my Custom GPT keeps producing bad output despite my Instructions?

Three diagnostic steps. First, check whether your Instructions are actually specific (not "match our voice" but "use sentences of 8 to 18 words"). Second, check whether your knowledge files contain off-brand examples that the GPT is accidentally imitating. Third, add one or two clear before-and-after worked examples in Instructions; this often unlocks dramatic improvement. If after three iteration cycles the GPT still underperforms, the job is probably too broad and should be split into two more specific GPTs.

Related reading

• Best AI marketing tools in Singapore 2026
• The AI marketing workflow that scales
• AI prompts for marketing teams