How I Write Prompts and Build Custom GPTs

✋Welcome to The CAS Cache, a newsletter designed to help accounting firms grow their CAS offerings in five minutes or less.

Disclaimer: Some links below support my writing of this newsletter, and some give you a deal.

Issue Sponsor 😎

Don’t outsource. Hire direct in the Philippines.

You can hire an accountant who works for you and only you. No middlemen. No ongoing fees. Just you and your team. Schedule a call with TeamUp now to learn how direct hire works.

Want to sponsor The CAS Cache? Reply to this email to get more details.

After the response rate to my last post on using AI for value pricing, I wanted to go a level deeper and talk about how I actually write prompts and build custom GPTs.

I build custom GPTs all the time now. Anything I find myself doing repeatedly, whether that is marketing, sales, pricing, content, or even creating prompts, starts to feel like a candidate for its own bot. Over time, that has become one of the biggest time-savers in how I work.

Two ideas in particular changed how I think about building them.

The first was learning to build a team into the prompt.

The second was using that same team-based approach to create a custom GPT whose only job is to help me write better instructions for future GPTs.

Build the workflow before the output

The biggest shift for me was realizing that most people prompt AI as if they are assigning one task to one person. Write this email. Summarize this article. Draft this proposal.

That can work, but it usually pushes the model straight to the answer. You get output quickly, but not always the best thinking behind it.

What I learned at a live AI seminar was to structure prompts more like a workflow. Instead of asking the model to act like one person, you build a sequence of roles into the instructions so it approaches the work from different angles before reaching the final output.

A good example is a marketing GPT I built. Instead of acting like one generic writer, I designed it to operate more like a small agency. It starts by analyzing audience psychology, motivations, pain points, fears, and desired outcomes. Then it builds the messaging framework. Then it writes. Then it reviews the work as a creative director would. Depending on the project, it can layer in SEO thinking or campaign strategy. Then it revises and gives me the final version.

That is still one model. It is not a true team of separate agents behind the scenes. But it creates a much better result because the model is being asked to think in sequence rather than jump straight to a draft.

That is actually one place I can improve the value pricing bot. It was useful, but I did not do as good a job building the team into it as I could have. There is still room to improve the workflow within it.

That has become one of the bigger lessons for me. Even when a custom GPT is helpful, you can usually improve it by being more intentional about the roles and sequence behind it.

Meta-Instructions GPT

The second unlock was even better. I started applying that same team-based idea to build a custom GPT whose sole purpose is to create meta-instructions. In other words, it helps me build better custom GPTs.

I cannot believe I had not thought of this earlier.

I spend a lot of time building bots and prompts. Once you do that often enough, you realize the hard part is not usually the idea. It is taking what is in your head and turning it into clear instructions, workflows, tone rules, exclusions, formatting requirements, and knowledge sources that the GPT can reliably use.

That is where a meta-instructions GPT becomes so helpful. Instead of starting from scratch every time, I can describe to the assistant what I want to build. What it should do, who it is for, what tone it should have, what knowledge it should use, and what I want it to avoid. Then the GPT helps me turn that into stronger underlying instructions.

Why I keep building custom GPTs

When I sent out the value pricing bot prompt, I realized it made sense to include directions on how to create a custom GPT, too. The reason I think they are so valuable is simple. If I am doing something repeatedly, I usually think about building a custom GPT for it.

That might be marketing. It might be sales. It might be writing proposals. It might be creating prompts for me. If I notice I am repeating the same thought process enough times, I start asking whether that process deserves its own reusable tool.

That is where the real leverage comes from.

You are not just saving time on one task. You are building something you can come back to over and over again.

And unlike one-off prompts, custom GPTs get better as you refine them. You can upload knowledge from subject matter experts. You can give them your own writing. You can feed them pricing material, frameworks, examples, and preferences. Then you can keep updating the instructions to include what you want more of and exclude what you do not want going forward.

That is a big reason they have become one of the best time-savers for me.

Where Projects fit in

I have also played around a bit with Projects, and I like a lot of what I see there.

I like the organization aspect. I like being able to keep related work grouped. I also like the ability to do deeper work inside that context.

But for me, Projects and custom GPTs solve different problems.

Projects are great for organizing work and keeping context together.

Custom GPTs are better suited to repetitive workflows.

That is the distinction I keep coming back to. If I want a reusable assistant for something I do repeatedly, I usually build a custom GPT. If I want a clean place to keep related work together, Projects are helpful.

Final takeaway

If you are just getting started with custom GPTs, do not make it more complicated than it needs to be.

Start with something you do repeatedly. Then ask yourself two questions.

What roles would make this output better?

What instructions would make this easier to reuse next time?

That is the shift that has mattered most for me.

Not just writing better prompts. Building better systems.

Thanks for reading, Luke Templin!