AI For Product Leaders: From Knowing To Doing

From pain points to practical systems: What Product leaders learned at our AI workshop

AI is rarely off the agenda at our Product roundtables, but the conversation can too easily stay theoretical. Our latest session was designed to change that.

We brought together a group of Product leaders for a hands-on workshop led by Maria Chilikov, former CPO at Beauty Pie and now CPO at AI startup Dastro.ai. Rather than discussing AI in the abstract, Maria walked the group through real examples she has built with advisory clients, and we spent the session building our own tool together.

In a pre-event survey, one of the most common recurring tasks the group flagged was writing weekly status updates for SLT and fielding the same team questions week after week. So that's where we started. Each attendee left with their own Custom GPT or Claude Project, a coaching partner that brings status updates to a consistent quality standard and critically, takes that task off the product leader's plate. Here's what the group learned.

What are Custom GPTs and Claude Projects?

Custom GPTs (in ChatGPT) and Projects (in Claude) let you build personalised AI models trained on your own instructions, context, and examples. Think of them as cloning your own standards, preferences, and ways of working, so the model can handle repeatable tasks to your specification, without you being in the loop every time.

Getting Started: Build for value, not novelty

The biggest trap when exploring these tools is getting distracted by what's interesting rather than what's actually useful. Maria's advice was clear: start with your highest-value problem, not your most exciting one.

• Focus on ROI first. Identify the task that takes up a disproportionate amount of your time every week and solve that. Flashy use cases are tempting, but the goal is to free up real hours for higher-value work.
• Start small and layer up. Break the problem into its smallest useful component. Aim for a workflow that gets you 70–80% of the way there, then build from there. You don't want to automate to 100% - keeping a human layer of approval and quality checking is the point (not a limitation).
• Add complexity in stages. Get one part working well first. Then add delivery automation, then data connections, then feedback loops. Building everything at once makes debugging almost impossible and frequently ends in something un-finished or un-usable.

Writing better prompts and instructions

How you set up your model determines how useful it is. The quality of your instructions matters far more than the sophistication of the tool.

• Treat the model like a new colleague. Give it enough context to do the job well, but don't overload it. Be explicit about: what you want it to do, who it's for, when it will be used, what good looks like, and what source of truth it should use.
• Start with instructions before adding a knowledge base. Get the prompt working well on its own first. Test it thoroughly. And only then layer in additional documents or context. Adding too much context too early can confuse the model and dilute the output.
• Don't overstuff context. Give the most important information, not all the information. Less is often more when it helps the model stay focused on the actual task.
• Make source-of-truth rules explicit. If certain documents or people should take precedence for different types of decisions, spell that out. Don't assume the model will infer it.

Evaluating and refining outputs

A polished-sounding output isn't the same as a good one. This is where human judgment remains essential.

• Good prompt design is about judgment, not just wording. The model can sound confident and still miss what actually matters for your business. Review outputs critically against your real standards and priorities.
• Be concrete about success criteria. Define what makes an output usable in pass/fail terms. What would make you send it? What would make you reject it? Vague scoring systems are much harder to act on.
• Use real examples to calibrate the model. Share examples of outputs you consider good and bad, then ask the model to refine its criteria based on what distinguishes them.
• Don't trust the first draft just because it sounds polished. Review outputs line by line and keep refining the prompt until what you're getting is genuinely useful and something that you would send.
• Feed corrections back in over time. Compare what the model produced against what was actually sent or used. That gap is your most valuable source of improvement, so use it to sharpen your instructions.

Building for your team, not just yourself

Some of the most valuable applications aren't about saving your own time, but about raising the standard across your team.

• Use the model as a coach, not just a writer. One of Maria's strongest points was to resist making it too easy to copy and paste. In the status update demo, the goal wasn't to write updates for the team, it was to coach them to write better ones themselves. Building in feedback loops and guidance makes the tool more valuable long-term and keeps the team's thinking sharp.
• Encode audience preferences. These tools can be personalised to a degree that most people don't take advantage of. You can build different models tailored to different stakeholders: who wants data first, who needs context first, what language lands with your CFO vs your CTO, what to include or exclude in a company-wide update, what nuances matter when you're presenting to the board. The personalisation opportunity here is significant and you can build all this into the instructions.
• Test with a trusted user first. Before rolling out a new prompt or tool more widely, test it with someone who will give you high-quality, honest feedback. It saves time and improves the output before it reaches a broader audience.
  
  

Working faster day to day

A few practical habits that the group found made a real difference:

• Use voice to work faster. Maria and the group strongly recommended speaking to AI models rather than only typing, especially when you're refining ideas. The back-and-forth is faster and often surfaces what you actually want more quickly than typing does. Wispr flow is a voice-to-text app many of the group use and rate.
• Iterate beyond the first session. These aren't set-and-forget tools. The best workflows improve over time as you feed in better examples, refine your instructions, and learn where the model's defaults don't match your standards.

If you'd like to learn more about the demo and other usecases being used by Product leaders, reach out to Maria or Louise, they'd both be happy to share more detail about the session and how to apply it to your own team.

Burns Sheehan are a technology recruitment agency, working with Product leaders across the technology sector. We host regular roundtables and events for our community — if you'd like to be included in future sessions, get in touch.

📧 [email protected]