Bowst Team
While the Bowst team is busy building next-gen digital solutions, I’m here as a student intern, working alongside their developers to deepen their AI skills – and mine. One of the most eye-opening experiences so far has been diving into Prompt Design with Vertex AI.
If you’ve ever wondered why AI sometimes nails an answer and other times goes completely off the rails, the answer usually lies in prompt engineering.
What I Worked On
This lab focused on how to design prompts that get large language models (LLMs) to respond exactly how you want. I explored how prompt style – from concise to vague, single-task to multitask – significantly impacts AI responses.
I experimented with prompt examples like:
- Vague + Multi-task:
“What is a good name for a website and what should I have for dinner?”
Result: The AI struggled with the split focus.
- Concise + Single-task:
“What is a good name for a website for a company that does ‘xyz’ for the target audience ‘abc’.”
Result: Clearer, more relevant answer.
I also played with system instructions (SIs), which are like pre-prompts that set the tone and rules for the model. For example:
“You are an AI chatbot for a travel website. Answer only fact-based questions. If the question is opinion-based, say: ‘Sorry, I can’t answer that.’”
These kinds of guardrails reduce hallucinations (aka when the model makes things up), which is a big deal if you’re trying to build reliable AI apps.
Tips I learned from Prompt Design
- Be clear, specific, and focus on one task per prompt
- Avoid combining unrelated questions – AI gets confused
- Examples improve quality dramatically
- System instructions help keep answers accurate and on-task
- For more consistent outputs, reframe generative tasks into classification tasks
Why This Matters
Prompt design isn’t a “cool trick”. It’s essential. If you want generative AI to be useful (and not frustrating), good prompting makes all the difference. Whether you’re building a chatbot or generating content, understanding how to guide the model makes it a powerful tool rather than a wildcard.
My Hands-on With Vertex AI Studio
Beyond text, I also got to explore Vertex AI’s Media Studio, where I generated images using Imagen and videos using Veo – all from text prompts!
With the Media Studio UI, I could control temperature (creativity vs. consistency), number of outputs, and other parameters. It really opened my eyes to how AI can transform a few words into high-quality media content.
Coding With Vertex AI: The Workbench Experience
Next, I stepped into the Vertex AI Workbench, using Jupyter Notebooks to run generative AI models programmatically. I:
- Connected to the GenAI API
- Tweaked system instructions and model parameters
- Ran batch predictions
- Practiced function calling and context caching for smoother, more efficient responses
This part gave me a deeper understanding of how to move from the UI to a code-based workflow, which is essential if you’re looking to scale or automate your AI applications.
The Prompt Design Challenge
To wrap things up, I took on a creative challenge using the Gemini 2.0 Flash model in Vertex AI:
Goal: Generate product descriptions, ad taglines, and poetic phrases based on images.
Tools: customizable prompts, keyword control, creativity sliders, and output length settings.
Insights: even small prompt tweaks can dramatically change tone and focus.
Being able to move seamlessly between Studio and Workbench made it easy to test and refine prompts, then scale them up for real-world applications like content creation and marketing.
Final Thoughts
Prompt design might sound like a niche skill, but it’s actually a cornerstone of effective AI interaction. What I learned in Vertex AI:
- Better prompts = better AI results
- System instructions are your secret weapon against hallucinations
- Gen AI isn’t magic – its logic, creativity, and clarity working together
Whether you’re designing chatbots, generating media, or automating tasks, a little prompt engineering goes a long way. And for someone like me – learning the ropes from the inside – it’s a game changer.
Big thanks to the team at Bowst for pushing me to learn, experiment, and now share.
