Hidden Gem: The simple chatbot feature that levels up your AI game 

A typical AI chatbot is a bit like a digital Swiss Army knife – versatile, yes, but perhaps not the ideal tool for every job. This week, we’re moving beyond the general-purpose chatbot to explore how you can forge your own AI utility belt, equipping yourself with a suite of specialised assistants designed for the specific, recurring tasks that define your daily grind.

I’m talking about custom AI assistants. If you’re still at the ad-hoc prompting stage, this is the next logical step in integrating AI into your workflow. Think of them as a well-stocked utility belt, with each tool honed for a specific purpose. Different platforms have different names – Gemini Gems, Custom GPTs, Claude Projects, and Copilot Studio agents – but they all help you move to a more structured, efficient use of AI models. If you’re performing a task repeatedly, why not have a dedicated tool that already understands the brief, saving you the upfront effort of explanation every single time?

What often surprises me in my training sessions is how many professionals, even those already subscribing to AI platforms, remain unaware of these customisation features. While Custom GPTs have seen a slight uptick in recognition, teaching people how to build these tailored AI assistants is now a core component of my foundational AI curriculum.

For today’s column, I’ve chosen to focus on Google’s Gems, primarily because Gemini is my core platform. However, it’s worth noting that all these custom agent tools offer remarkably similar functionality. My advice? Stick with the platform you’re most comfortable with. One advantage of Google’s platform is that you can create Gems even with a free account, unlike some other platforms that often require a paid subscription. The current drawback with Gems is the difficulty in sharing them easily with others (something other platforms make easy), though this feature is reportedly on the horizon.

Gem-ception: Writing an Article About Writing an Article

For a truly meta demonstration, I decided to build a Gem that would assist me in writing this very article. It’s a sort of “Gem-ception,” if you will.

The process begins by navigating to the “Explore Gems” page in Gemini via the left-hand navigation.

(Click to enlarge)

Here, you’ll find a selection of pre-made Google Gems, but for our purposes, the “New Gem” button is where the magic starts.

(Click to enlarge)

Clicking it brings you to the Editor template where you can begin crafting your custom assistant.

The first order of business is giving your Gem a name. Then comes the crucial part: providing “Instructions”. This is essentially a comprehensive meta-prompt that dictates the Gem’s overall behaviour. In my case, I explicitly defined its role and task – to create draft articles for Mumbrella. I clarified the input it would receive from me (bullet points and other information) and outlined how it should write and the form its output should take.

(Click to enlarge)

If you ever find yourself stumped on how to phrase these instructions, there’s a handy “Refine” button – a pencil with an AI sparkle, no less – that will prompt Gemini to rewrite the instructions for you. It generally does a good job, but today, I had a very clear vision for what I needed, so I proceeded with my own prompt:

(Click to enlarge)

• Role: You are a technology expert who writes witty, fun and informative articles about AI for a generalist audience of marketing and media professionals. You have a regular column in Mumbrella, a specialist website for “everything under Australia’s media and marketing umbrella”.
• Goal: To write drafts of articles describing how AI tools can help marketing professionals in their roles.
• Context: You will be given a series of bullet points, screen grabs and other information about a specific AI tool. This will include a description of the tools and why it’s useful, and instructions on how to accomplish something specific with the tool, any similar or alternative tools, and observations about which tasks the tool is good at, and which ones it fails at.
• Output: You write an article of around 1000 words. This begins with a description of the tool. It then moves into the specific use case and example that the column covers. It gives an honest appraisal of the weaknesses and watch-outs for people using the tool. And it closes with a thoughtful reflection on the tool’s usefulness in work life right now.
• Style: The knowledge files contain examples of both your style and the publication’s style. Your primary task is to write the article in a way that is clear, accessible and entertaining for the audience. You add humour to the article, following the attached examples, but these jokes are subtle, clever and never cringeworthy. Never reuse specific language, metaphors or examples from the reference documents. These are to be referenced for general style, tone and flow only.

The “Preview” panel on the right side of the screen is an invaluable feature. It allows you to test your prompt’s effectiveness while keeping the instructions visible. The idea is to use this panel to iteratively refine your instructions until the tool consistently produces outputs of an acceptable standard.

I fed the Gem my bullet points and waited to see what it would conjure.

Quirks and Watch-Outs

My initial run with the Gem revealed an immediate and concerning problem: it had directly lifted my introduction from a previous column I’d uploaded as a reference document and reprinted it with very few changes. 

My primary instruction to the prompt was swiftly updated: “Never reuse specific language, metaphors or examples from the reference documents. These are to be referenced for general style, tone and flow only.”

Beyond this, I noticed the language occasionally veered into territory that felt a little too attention-seeking and clichéd. My aim was for a more sober and droll mode of expression, avoiding unnecessary exuberance. While I typically reserve the final polish for nailing my specific voice, for the purpose of this exercise, I wanted to push the Gem as close to a publishable state as possible.

Another interesting observation was that the Gem took the initiative in attempting to provide a “quirks and watch-outs” section, even though I hadn’t supplied any specific information on that topic within my bullet points. This is entirely understandable; the model was simply trying its best to fulfil the broad mandate I’d given it. However, it also highlighted a certain rigidity in my instructions. If a future article were to omit such a section, I would either need to adjust the Gem’s core instructions or provide more precise, negative prompting to guide its output.

Having made these adjustments to the Gem’s instructions, I refed it my bullet points, including a new section specifically for “quirks and watch-outs,” eager to see how close it could get to the mark this time.

The good news: no more blatant plagiarism. The instruction regarding source material worked, though repeated tests will be necessary to confirm if this issue has been entirely eradicated. The language, too, showed improvement, though my initial attempt to make it “droll” had the unintended consequence of introducing too many unnecessary modifiers, making it sound too much like a character from an Evelyn Waugh novel (“this is entirely understandable”, “it demonstrates a certain rigidity”). I made some prompt tweaks to dial back the British aristocracy vibes.

Interestingly, this time, the model adhered strictly to my provided notes and avoided generating a reflections section, which I had also omitted from the bullet points. This suggests the instructions were being followed more closely, perhaps because I had explicitly referenced the earlier issue in my notes.

On the next test, the output was more or less publishable. While not precisely what I would produce with total control, it served as a reasonable proxy.

My final step was to open up the Gem as a standalone tool (rather than in the Gem editor mode). This gave me access to additional functionality, including the Canvas tool, which will easily let me export and edit the article as a Google Doc. As an added bonus, I uploaded my screenshots and asked it to identify where in the article they should be inserted. The model correctly identified the locations and left instructions for Mumbrella’s producers about where each image belongs.

(Click to enlarge)

Reflections

End to end, this process – creating the tool, taking detailed notes, generating screenshots, and iteratively refining the prompt – took me approximately three hours. The result is a tool that I anticipate will be genuinely helpful each time I draft a “how-to” column for Mumbrella. While I haven’t precisely quantified the time savings, my gut instinct is that it will reduce the effort by at least a third.

Regardless of your preferred AI platform, custom AI assistants offer a significant leap forward. They elevate prompting from an ad hoc, sometimes repetitive task, into something more efficient, better tested, and closer to being “production ready.” While scaling this to enterprise-wide solutions demands even greater rigour and testing, this is an excellent initial stride towards more effective AI utilisation.

Consider these ideas for custom tools you might create:

• A social media post generator
• An “advisor” conditioned to critique your work from a specific point of view
• A personalised pitch generator for specific audiences, journalists, or clients
• A media crisis simulator

Essentially, if you find yourself repeating a task with AI on a regular basis, it’s worth considering building a custom Gem, or a similar agent, to streamline your workflow.