AI for Accountants Vol. 3 - Prompting 201
Welcome to issue #006 of New Age Accounting — Vol. #3 of AI for Accountants. Prompting 201.
Thanks for coming back for Vol. 3 of AI for Accountants!
Quick recap: Vol. 1 covered AI & LLMs — what they are, which ones exist, and how to choose one. Vol. 2 introduced prompting and the CO-STAR framework — the foundation of getting outputs that are actually usable. If you missed either, go back and start there (using links provided) as this series builds week by week.
Now, before we go any further — did you use the CO-STAR framework to enhance your prompting? Because if you did, one of two things likely happened. Either it worked better than expected, and you’re starting to see what all the hype around AI is. Or, it worked okay, the output was close, but you had to rewrite more than anticipated.
If it’s the latter, that’s perfect — that’s exactly what Vol. 3 is for.
Why good prompts still fall short.
Here’s something I’ve noticed working with these tools daily: even a well-written prompt can fall short if the LLM doesn’t have the right frame for the task.
What does that even mean?
Think about how you’d provide context to a first year associate or new accountant on the team. You wouldn’t just hand them a task and walk away. You’d walk them through their specific role in the process, what standard you’d like set, what a great outcome looks like, and potentially provide examples of something in the past (SALY, if you know, you know).
That context — or the frame around the task — is really what separates a good output from a great one.
Good outputs come from people giving the LLM the task. Great outputs come from giving the LLM both the task and the frame.
Let’s get into the meat of it. Here are four techniques to creating a strong frame to immediately improve your outputs. None of them are complex but all of them provide real value to your prompting skills.
Technique 1 : Establish a role
This is going to sound strange. But, before you describe a task, tell the LLM who it is. Weird, I know. However, when you set the stage for the LLM, it activates a specific body of knowledge and strives to achieve in line with that specific standard.
Here’s an example without a role:
“Review this journal entry and flag any issues”
Here’s an example with a role:
“You are a senior auditor reviewing journal entries for a public company. Review this journal entry and flag anything that would raise a question in an audit. Be specific about what you’d like to see as audit support.”
Same task. However, the second prompt will produce an output that is materially more useful because the LLM is approaching it from a specific lens, instead of a general one. This works for controllers, auditors, analysts, and more.
Try it on your next prompt. The difference will be immediate.
Technique 2 : Provide an example
Just like a first year or someone new to the company, if you want a specific output format, provide an example to the LLM.
Instead of describing the format you’d like to see — which is much more difficult than it sounds — just provide an example of something similar for it to work from. It’s as simple as uploading the file and saying, “use the uploaded document as a template”.
Here’s what that looks like in practice. Say you want the LLM to write a board-ready variance commentary. Instead of writing “format it like a board report with bullet points and an executive summary” — just upload last month’s commentary and say “match this format exactly.“ The LLM picks up the tone, the structure, the level of detail, and the language all at once.
Once provided, the LLM then has a template, tone, task and more to work from. The output will match your format far more closely than if you tried to describe it from scratch.
An important thing to note, once you’ve created this task, make sure you save the prompt with the example baked in. I typically rename my chats and pin them, if I’m going to be using them later on. If you do this, you will be able to use the same chat every month and it saves more time and tokens because you aren’t reprocessing the prompt end to end.
Technique 3 : Ask to be asked
Hands down, this is one of the most underused techniques. Before you write a complex prompt, try this:
“I need help with [insert task you’re working on]. Before you start working, ask me any clarifying questions that would help you give me a better prompt.”
Simple, right?
Two things happen when you do this.
First, the LLM finds gaps and weak points in your prompt you likely didn’t know were there. It’ll ask about your audience, your format preference, the context behind the numbers, the tone, and more. It takes the CO-STAR framework to the next level. Most of which, are things you would have forgotten to include.
Second, it teaches you how to prompt better over time. The questions AI asks you are the exact things you should be thinking about and including in your prompts by default. After a few rounds of this, you’ll start anticipating these questions automatically.
Any time I’m prompting something new for the first time, I do this. It almost always produces a better first output than going straight to the task.
Technique 4 : Refine, don’t restart
This one is a mindset shift as much as a technique.
When the output isn’t quite right, most people start over. New prompt, from scratch, trying to get it right in one shot.
Don’t do that.
Remember what Vol. 1 said: you’re not Googling. You’re having a conversation. That means you can — and should — push back, redirect, and refine within the same conversation.
Here’s what that looks like:
First output comes back. It’s okay but too formal for your CFO’s style.
Instead of starting over, say:
“This was a great foundation. The tone is too formal. My CFO prefers plain language, short sentences, no jargon. Rewrite the second paragraph with that in mind.”
Each refinement builds on the last. By the third or fourth exchange, you have something significantly better than anything a single prompt would have produced.
The conversation is the tool. Use it.
Bringing it all together
Alright. We’ve gone through a lot in the last few weeks. Here’s what a fully loaded prompt looks like using everything from Vol. 2 and Vol. 3 combined:
“You are a controller preparing the October close commentary for a CFO and board audience. Here’s an example of the format and tone we use: [paste example]. Using that same structure, write the variance commentary for the following results: [paste data]. Marketing came in $47K over budget driven by a pulled-forward campaign and an unplanned conference sponsorship. Revenue beat by $120K — two enterprise deals closed in the final week of the month. Before you start, ask me any clarifying questions that would improve the output.”
That prompt has a role, an example, real context, and an instruction to clarify before it starts. That’s about as strong as a prompt gets.
Try it this week on something from your actual close. See what comes back. Refine once. Save it.
The accountants putting these into practice right now are producing work that would have taken hours in a fraction of the time — and they’re getting better at it every week. That’s the compounding effect of learning this properly. Vol. 4 is next. Keep building.
Here are my questions to you:
Which of these four techniques are you most excited to try — and which one surprised you most? Click below and tell me. I read every response.
The purpose of New Age Accounting is simple: to empower accountants — at every level — to become builders, not bookkeepers. Whether you’re a staff accountant, a controller, or a CFO, there’s something here for you. Some topics will be high level, others will come with step-by-step guides, and some will include the exact prompts and tools you need to start building today.
If you’ve made it this far, you’re already thinking differently about this profession. Subscribe and come build with us.