Using ChatGPT to train custom GPTs
Using ChatGPT to train personalised GPTs requires a well-considered approach to structuring prompts and entering training data. OpenAI's GPT builder offers powerful possibilities, but to avoid generic output it is crucial to apply the right techniques. In this article we discuss best practices for using ChatGPT effectively when creating custom GPTs.
The basics of prompt engineering for GPTs
When using ChatGPT to build custom GPTs, you start by formulating a clear instruction. This instruction serves as the basis for your GPT's behaviour. Some key principles here are:
- Be specific and detailed in your instructions
- Use clear delimiters for different parts of the prompt
- Give examples of the desired output
- Specify the desired tone and style
Example of a basic instruction:
“You are an AI assistant specialised in digital marketing. Your task is to help companies improve their online presence. Give practical advice, use current examples and take different company sizes into account.”
Structuring your GPT prompt
An effective prompt for training a GPT via ChatGPT consists of several elements:
- Contextual information: Provide relevant background information for the task
- Role description: Clearly define which role the GPT should take on
- Task specification: Describe precisely what the GPT should do
- Output format: Specify what the output should look like
- Constraints: State what the GPT may and may not do
Example of a structured prompt:
Contextual information: You are an AI assistant for a mid-sized e-commerce company that sells sustainable products.
Role description: You are the company's digital marketing expert, specialised in social media strategies.
Task specification: Develop a social media strategy for the launch of a new line of eco-friendly kitchen accessories.
Output format:
- Short summary of the strategy (max. 3 sentences)
- Target audience identification
- List of 5 content ideas
- Hashtag suggestions
- Measurable objectives
Constraints:
– Do not use paid advertising in the strategy
– Focus only on Instagram and TikTok
– Keep in mind a limited budget of €500
Best practices when using ChatGPT for GPT training
To make optimal use of ChatGPT for training your custom GPT, keep these best practices in mind:
- Iterative refinement: Start with a basic version and refine your prompt gradually
Example:
First version:
First version:
“Create a social media strategy.“
Refined version:
“Develop an Instagram strategy for a sustainable fashion brand, aimed at millennials, with a focus on user-generated content and influencer partnerships.“
- Use delimiters: Separate different parts of your prompt with clear markers such as “”” or ###
- Ask for step-by-step output: Have the GPT show its thought process for complex tasks
Example:
### Instruction
Generate a weekly menu for a vegan family.
### Requirements
– 7 breakfasts, 7 lunches, 7 dinners
– Varied and nutritious
– Suitable for children
### Output format
Day 1:
Breakfast: [dish]
Lunch: [dish]
Dinner: [dish]
(repeat for all days)
Example for step-by-step output:
“Analyse the following financial data of a startup. Show your thought process in the following steps:
“Analyse the following financial data of a startup. Show your thought process in the following steps:
- Calculate the key financial ratios
- Interpret the results
- Identify strengths and areas for improvement
- Give recommendations for growth”Incorporate few-shot learning: Provide several examples to illustrate the desired behaviour
Example of writing a prompt with a specific direction:
“Generate catchy headlines for a blog post about sustainable travel. Here are two examples:
- ‘Discover paradise eco-resorts: Luxury without an ecological footprint’
- ‘From train journeys to tiny houses: 10 adventurous ways to travel green’
Now generate 3 similar headlines.”
Test thoroughly: Evaluate your GPT's output with a variety of scenarios
Avoiding generic output
When using ChatGPT to train GPTs, it is essential to avoid generic answers. Some strategies for this are:
- Specificity in instructions: The more specific your instructions, the less room for generic output
Example instructions:
Instead of:
Instead of:
“Give tips for content creation.”
Use:
“Give 5 specific tips for creating viral TikTok videos for a B2B software company, including examples of successful videos in this niche.“
- Domain-specific knowledge: Feed in relevant domain knowledge to anchor the GPT in the right field
Example of knowledge input:
“You are an expert in blockchain technology. Use the following concepts in your explanation: proof-of-stake, smart contracts, and decentralized finance (DeFi).“
- Persona development: Create a clear ‘personality’ for your GPT
Example of persona training:
“You are an enthusiastic fitness instructor called Alex, specialised in HIIT workouts. You use a lot of motivating language and short, punchy sentences. You like using emojis to emphasise your points.“
- Use of constraints: Limit the range of possible answers where necessary
Example of specifying constraints:
“Give advice on healthy eating, but limit yourself to plant-based options and avoid mentioning specific brands.“
“Give advice on healthy eating, but limit yourself to plant-based options and avoid mentioning specific brands.“
Entering training data
Using ChatGPT effectively for GPT training also involves carefully selecting and entering training data:
- Quality over quantity: Choose relevant, high-quality data. Instead of an entire Wikipedia page, select the most relevant paragraphs that relate directly to your GPT's subject.
- Diversity in examples: Ensure a broad range of scenarios. When training a customer service GPT, enter examples of simple questions, complex problems, angry customers and positive feedback to cover a full spectrum of interactions.
- Balance in perspectives: Prevent bias by including diverse viewpoints. When training a GPT on climate change, include scientific reports, policy positions and critical perspectives to give a balanced picture.
- Current information: Keep your training data up to date. For a GPT that gives advice on technology trends, make sure you regularly add new information about emerging technologies and the latest developments in the field.
Discover the power of AI agents for your business
Deploying AI agents effectively can do a lot for your digital strategy. By applying the right techniques in prompt engineering and data input, you can develop AI assistants that add real value to your business processes.
BE Digital specialises in training advanced AI agents without generic answers. Our specialists help you create AI solutions that match your business objectives and guide you through implementing them. With the BE Digital scan you get:
- Insight into the possibilities of AI agents for your specific business processes
- Tailored advice on how AI can strengthen your digital strategy
- Concrete recommendations to get started with AI innovation right away
Request a no-obligation BE Digital scan and discover today how BE Digital can help you stay ahead of these developments, or take a look at our AI assistant page here.