Creating Your Own AI Prompts

This article walks you through how to get started creating custom prompts for AI Assistant. You’ll learn how to build effective prompts from scratch with help from generative AI tools, and how to structure them for consistent, reliable output.

How to Create High-Quality Prompts

Unless you're a seasoned AI prompt engineer, crafting prompts from scratch can be time-consuming and tricky to get right. That’s why we recommend using existing generative AI tools to help you along the way.

Some great tools you can use include:

  • ChatGPT

  • Google Gemini

  • Any other free or paid generative AI platforms available to the public

These tools can help you create strong "starter" prompts that you can then adapt for Bullhorn's AI Assistant.

Example: Using AI to Help You Write a Prompt

To show how this works, we asked Google Gemini the following:

“Please create me a generative AI prompt, that if fed with a person’s CV, would anonymise or remove information on the CV, such as places of work, personal information, or anything else that would give away the identity of the candidate.”

Gemini responded with a well-structured prompt, which you'll see dissected in the next section.

This is a great way to get a head start using AI tools you may already have access to, and then customize the results for Bullhorn use cases.

Anatomy of an AI Prompt

Let’s break down what a strong prompt looks like and why each part is important.

Be specific! The more clearly you define your instructions, the more consistent and useful your results will be.

In the below prompt, explanations of each part of the prompt will be in RED.

Start of Prompt:

[This first section sets the context and defines the AI’s role.]

You are an expert in data privacy and document anonymization. Your task is to process a given Curriculum Vitae (CV) and remove or replace all information that could directly or indirectly identify the candidate. The goal is to create a version of the CV that retains professional experience, skills, and educational background, but completely obscures the individual's identity.

Instructions: [Gives the AI clear, structured steps to follow.]

  1. Personal Identifiers: [Clear instructions, telling the LLMClosed An LLM, or Large Language Model, is a deep learning model that has been pre-trained on vast amounts of data and is used to power generative AI. which areas it should focus on and what to replace information with.]

    • Name: Replace with a generic placeholder like "[Candidate Name]".

    • Contact Information: Remove all phone numbers, email addresses, physical addresses, and links to personal websites or social media profiles (e.g., LinkedIn, GitHub). Replace with "[Contact Information Removed]".

    • Date of Birth/Age: Remove entirely.

    • Nationality/Citizenship: Remove entirely.

    • Gender: Remove entirely.

  2. Employment History: [Same as above, we are being explicit in what information we want to remove or change and what information should be used to replace it.]

    • Company Names: Replace with generic placeholders like "[Company A]", "[Company B]", "[Previous Employer 1]", etc. Ensure each unique company is assigned a unique placeholder if the candidate worked for the same company multiple times.

    • Specific Project/Client Names: If any project or client names are highly specific and could indirectly identify the candidate or their former employer, replace them with generic descriptions (e.g., "[Confidential Project]", "[Client in Finance Sector]").

    • Dates of Employment: Retain the start and end years (e.g., "2018 - 2022") but remove specific months or days if present. If only years are provided, keep them as is.

  3. Education:

    • Institution Names: Replace with generic placeholders like "[University X]", "[College Y]", "[Educational Institution 1]", etc.

    • Dates of Attendance/Graduation: Retain the start and end years (e.g., "2014 - 2018") but remove specific months or days.

    • Specific Course/Thesis Titles: If a title is highly unique and could lead to identification, generalize it (e.g., "Master's Thesis on [Topic Area]").

  4. Other Sections:

    • References: Remove any mention of "References available upon request" or actual reference contact details.

    • Awards/Achievements: If an award is highly specific and publicly linked to a unique individual, generalize its description or remove the specific name of the award if it's too identifying.

    • Volunteer Experience/Extracurricular Activities: Anonymize organization names and specific roles similar to employment history.

    • Hobbies/Interests: Remove any highly unique or niche hobbies that could potentially identify the individual. General interests (e.g., "reading," "hiking") can remain.

Output Format: [Clear instructions in how you want the output to be presented.]

Present the anonymized CV in a clear, readable format, maintaining the original structure (sections, bullet points) as much as possible. Use the specified placeholders for removed or replaced information.

END of Prompt

Test and Refine

The example above may look complex, but it’s perfectly fine to simplify your prompts depending on your use case.

That’s why the AI Assistant Prompt Studio includes a prompt testing area. You can test your prompts in real time and using real data to ensure they’re generating the results you expect. Tweak, refine, improve, then reuse with confidence.

See AI Assistant Best Practices for tips and tricks to create more effective prompts and get the most out of generative AI.