Prompts for AI Tools Don't Matter Anymore. 3 Things That Change When You Brief Claude Like an FP&A Analyst

Full transcript

Lightly edited for readability. The audio is the original.

0:00The agent that ran a 15-prompt job while I got coffee

I spent 2 years getting good at prompting AI. The perfect wording, the magic phrases, the whole "act like a CFO with 20 years of FP&A experience" setup. I had a document with a couple hundred prompts in it, and I was pretty proud of it.

Then last month, I watched an AI agent take a job I would have broken into 15 separate prompts and just run it, start to finish. It opened the files, did the work, checked itself, and handed me the result, while I got a coffee. And it hit me, standing there at that counter, that the skill I'd spent 2 years sharpening was already halfway to useless. The perfect prompt wasn't the thing anymore.

Today, I want to show you the skill that replaced it. And here's the good part. It's a skill every decent finance manager already has. You've just been using it on people instead of on the machine. Let me show you.

Welcome to Excel at AI in Finance, the show for finance pros who want to use AI to get home sooner and be the one their company can't afford to lose. I'm Mike, I run F9 Finance, and I'm a senior finance leader working in the trenches. Each week, I share the AI tools, workflows, and conversations that get you where you want to be. All right, let's get into it.

1:15Dana had a couple hundred saved prompts and still lost her afternoons

Let me make this real with a member of the Finance AI Lab. I'll call her Dana, and I've changed the details that don't matter. Dana's a senior analyst headed for assistant controller at a mid-sized company, and when she joined the Lab, she was already the AI person on her team. You know the one.

She had the saved prompt library. A couple hundred prompts in a doc, sorted by task, tuned within an inch of their lives. If you needed the exact wording to get the model to write a clean variance narrative, you went to Dana.

The sad thing is she was still drowning, and that's the part that got me. Because here's what her month-end actually looked like. She'd sit down to reconcile a couple of accounts, and she'd run it like a relay. Prompt one, get the matching logic. Copy that out.

Prompt two, pull the exceptions, reformat it by hand. Prompt three, write the summary. Six or seven prompts, a pile of copy and paste in between, every one of them polished. And it still ate her whole afternoon.

She'd automated the typing and kept all the actual work. And she couldn't see it, because from the inside, it felt like using AI. She was using it. She just had it doing the fastest, easiest slivers of the job, the parts that were never the problem, while she personally carried every handoff in between. The tool was saving her seconds and costing her the afternoon.

2:37The one plain instruction that replaced the relay

Then she hit a roadmap in the Lab on working with agents. And the whole idea was one line. Stop writing prompts. Start writing a brief.

And you know what? She almost skipped it. She was good at prompting, and this felt like a step backward. But she tried it on that same reconciliation. And instead of the relay, she wrote one plain instruction.

Here are the two exports. Match them on invoice number. Flag anything unmatched over $5,000. And show me the rows you couldn't match, with the reason.

That was it. No magic words, no "act like a controller." If anything, it was a worse prompt than the ones in her library. Just a clear job. She hit go and went to get a coffee.

And when she came back, it had done the whole thing. Opened both files, worked the match, built the exceptions list, and handed her a first-pass write-up of what didn't tie and why. The afternoon job took about 10 minutes.

Now here's the part that stuck with me, because Dana said it better than I could. She said the thing that finally worked wasn't a new AI skill at all. It was the exact same thing she does every single week when she hands a task to her junior analyst.

Here's the file. Here's what I need. Here's what to flag. Come find me if you get stuck.

She'd been doing that with people for years. She just never once thought to do it with the machine. 2 years of collecting the perfect prompts. And the thing that finally changed her month-end was a skill she already had before she ever opened an AI tool.

That's the whole episode sitting in one story. So let me back up and show you the three things that change the day you stop writing prompts and start briefing the work.

4:19Why the tools stopped answering and started doing

So here's the shift underneath Dana's story. For about 2 years, the whole pitch on AI was learn to prompt. Get the wording right, and the model will do almost anything. And that advice made sense, because the tools only answered.

You typed a request, it wrote something back. And then you did the next step. And the one after that.

You were the engine moving the work from one prompt to the next. The prompt was the whole interaction. So of course, the prompt felt like the skill.

But the tools stopped just answering, they started doing. Claude, Copilot, the agent mode showing up right inside Excel now. You give one of these a job and it'll make a plan, open your files, run the steps, use other tools, and hand you back a finished result.

And the second a tool can run the whole job instead of one slice of it, the skill flips. It stops being about how you word one request. It becomes about how you hand off the whole thing.

And this isn't a someday thing. It's already sitting in the tools on your desk. The spreadsheet you live in has an agent mode now. The chat tool you already pay for will go open a folder, read 10 files, and come back with an answer.

The capability showed up quietly inside the tools you already use, which is a big part of why most finance pros haven't changed how they work yet. The tool leveled up, but the habit didn't.

5:44The gap between the people who notice and the people who don't

And the gap between the people who notice and the people who don't is going to get wide, fast. The finance pro still hand-feeding prompts is working at one speed. The one briefing agents is working at another. Same team, same tools, completely different output.

And that's the whole reason this episode exists. So let me give you three things that change the day you make that move. Dana lived all three in one reconciliation.

7:03Shift 1: the wording stops being the skill

The first thing that changes is that the wording stops being the skill. For 2 years, the craft was wordcraft. The exact phrasing, the roleplay setup: "You are a senior controller with 20 years of experience." The secret 10-part structure somebody sold you in a thread with 100,000 likes.

And with the old chat-and-answer tools, that stuff genuinely moved the needle. Because the model got one shot, and the words were the only thing it had to go on. With an agent, that barely matters. And I want you to really sit with that, because it's the hardest one to swallow for anybody who got good at prompting.

The magic words don't do much anymore. What does the work now is the brief: the job, the inputs, the limits, and what done actually looks like when it's finished.

Go back to Dana for a second. Her old prompts were genuinely good, tuned, worded, sorted in a library, and they had her running a seven-step relay every month-end.

Her new instruction was plain to the point of boring. Here are the two files. Match on invoice number. Flag anything over 5 grand. Show me what didn't tie.

There was no craft in that sentence at all. And that's the one that ran her whole afternoon in 10 minutes.

8:18The tell you can run tomorrow ("build me a variance analysis")

Here's the tell, and you can run it tomorrow. Take the ask that flops as a one-shot prompt: "Build me a variance analysis." Type that, and you get something generic and hollow back, because you didn't actually tell it anything.

Now, hand that same ask to an agent as a brief. Here's the file, here's last month's format. Walk every line that moved more than 10%. Pull the driver from the notes tab. Flag anything you're unsure about.

Same model, same data, but now it's useful. The difference was never the wording. It was whether you spent the 60 seconds to scope the job before you hit go.

And most finance pros skip that 60 seconds, and I get why. For 2 years, the whole internet told them the job was the prompt. Nobody told them the prompt was the easy part, and the thinking was the job. So they polished sentences and wondered why the machine kept handing back mush.

And it shows up everywhere in finance, not just reconciliations. The board commentary you rewrite every quarter, the forecast refresh, the ad hoc the CFO drops on your desk at 4 o'clock. Every one of those is a job you could brief instead of a prompt. And the plain, well-scoped version beats the clever one every time.

I've stopped being impressed by anyone's prompt. I get impressed by how clearly someone can describe the job. The fix in one line: stop wordsmithing the prompt. Write the brief you'd hand a new analyst on their first day.

9:47Shift 2: from one step to the whole job

The second thing that changes is where you sit in the work. With prompting, you're the glue. You take the job, you break it into steps in your own head, and you feed them in one at a time. You carry the output of step one into step two. You do the joins between the pieces.

The AI does little bursts of work, and you're the one stitching them into an actual finished thing. That was Dana's relay. Copy, reformat, prompt again. She was the conveyor belt running between all those little bursts.

With an agent, you hand over the whole job and let it run the steps itself. It makes the plan. It opens the files. It calls the other tools. It reads the tab, does the match, builds the list, drafts the summary.

Your skill moves up a level. It becomes cutting the job into a shape the agent can follow and setting the guardrails around it, not doing the handoffs yourself.

Month-end reconciliation is the clean example, because it's exactly what Dana ran. The old way, you prompt for the match logic, copy it out, prompt for the exceptions, format them, prompt a third time for the write-up. Five, six, seven touches. The agent way, you give it the two files, the matching rule and the threshold, and it produces the exception list and the first-pass narrative in one go, while you're not even at your desk.

11:06What this does to a quarterly reporting pack

Scale it up, and it gets obvious. Think about a quarterly reporting pack. Pull the actuals, compare to plan, build the variance table, draft the commentary, format it to the template. That's a dozen steps across three files. And for years, that was a full day of you being the glue.

Brief it as one job with the template and the rules attached, and the agent runs the whole chain and hands you a draft to react to. Your day goes from building it to judging it. That's the shift in one deliverable.

Now here's the tell that you're ready for this, and it's a good one. If you can describe the job as a series of steps a brand-new hire could follow, the agent can follow it too. And if you can't, that's not the agent's problem. That's the actual work, and it was hiding under all that prompting the whole time.

Half the reason people can't get an agent to do their job is that they've never once written down what their job actually is, step by step. They just do it on muscle memory.

12:04A work instruction is what an agent runs on

So the move here is almost old-fashioned. It's the thing a good manager does when they delegate. You don't hand your analyst a vague wish and hope. You tell them the sequence.

First pull this, then match it against that. Here's the rule for the weird cases. Here's what to do when you hit one you can't solve. That's not a prompt, that's a work instruction. And a work instruction is what an agent runs on.

And here's what that turns you into, which I think is the real point. You stop being the person doing the 15 little tasks. And you start being the person who runs the work. That's a promotion in how you spend your day. And you don't need anybody's approval to take it.

The analyst briefing three agents through three jobs at once is doing manager-level work, whether or not the title has caught up yet. That's the seat AI opens up, and very few are sitting in it.

The fix here comes down to one line. If you'd have to explain it to a new analyst, write that explanation down. That is the prompt now.

13:04Shift 3: from eyeballing the answer to checking the trail

The third thing that changes is how you check the work. And this is the one that bites finance pros the hardest, so stay with me. With a single prompt, checking is easy. You get one answer, you read it, you can eyeball whether it's right. One output, one look.

But an agent doesn't give you one output. It runs a dozen steps and hands you the end of the chain. And you cannot eyeball a dozen steps you never saw.

You don't know what it opened. You don't know what it assumed when it hit something ambiguous. You don't know which of those steps quietly went sideways.

13:36The reconciliation that balanced to the penny and was wrong

Let me give you the version that should worry you a little, because I've watched it happen. An agent reconciles two ledgers and hands back a clean, confident summary. Everything ties. Looks perfect. And buried three steps back, it hit 40 rows that didn't fit, made a quiet assumption to force them to balance, but never mentioned it.

The output looked finished. It balanced to the penny. And it was wrong in a way you'd only catch if you went and looked at the trail. That's the finance version of a car that drives beautifully with a crack in the frame.

14:07The verification tax, applied to agents

This connects straight to something I hit a couple of episodes back, the verification tax. Everything AI hands you has to be checked before you can trust it. And with an agent, the check isn't reading one answer. It's auditing a process you didn't watch.

So the whole skill becomes scoping the job so it's checkable in the first place. You build the receipts into the brief.

14:30What checking the trail looks like

So what does checking the trail look like? You don't rerun the whole thing. You spot-check the seams. You ask it right there in the brief for the three or four things you'd want from a junior anyway.

The rows it matched and the ones it couldn't. The assumptions it made when the data got messy. The count that should tie, tied out loud. Give me those, and a 2-minute read tells you whether to trust the 20 minutes of work underneath. Skip them and you're back to redoing it just to be sure.

That's the split between the finance pros who get burned by agents and the ones who get their hours back. The ones who get burned read the last line. It looks great, and they forward it. The ones who win told the agent up front: cite the rows you matched, flag anything you were unsure about, and tell me exactly what you skipped and why.

Same agent, but one of them gets a black box. The other gets a result with its work shown that they can sign off in a glance. And that one habit, asking for the work to be shown, is the thing that keeps your name off a mistake the machine made on your behalf.

The fix? Brief the agent to show its work and flag its own shaky calls. Otherwise, you're not delegating. You're signing your name to a black box.

15:40The skill you already had before you opened an AI tool

The thing I really want to emphasize is that you already know how to do this. It's the same move you make with a good analyst. Name the job, hand it over, say what done looks like, check what comes back.

2 years of prompt tips talked you out of a skill you already had. Agents just handed you the analyst.

And this matters way past your next reconciliation. The finance pros who learn to brief the work the way you'd brief a sharp analyst become the person their team trusts the day the tools change again. And they will change again. That's the person everyone else turns to.

16:153 questions to run before you hand a job to an agent

Okay, so what do you do with this before your next close? I want to give you three questions. Run them before you hand any job to an agent.

Question one: could I hand this to a brand-new analyst with exactly what I just typed? If the brief isn't clear enough for a person, it isn't clear enough for the agent either. A new hire would come back with questions. The agent won't, it'll just guess. So answer the questions before they get asked, right there in the brief.

Question two: did I tell it what done looks like and what to flag? An agent with no target and no guardrails hands you something confident, plausible, and quietly wrong. And you own it, with your name on it.

Say what finished means. Say what to raise its hand about. That one move heads off most of the messes.

Question three: can I check this in a glance, or did I ask it to show the trail? If you can't verify what it did, you're not delegating, you're gambling. So build the receipts into the ask. Show your work, cite the rows, flag what you skipped.

And here's the hard version. If you're briefing an agent the way you'd fire off a quick prompt, lazy, vague, half a sentence, you're not saving yourself time. You're moving the mistakes downstream, to a place where they're harder to catch and still trace straight back to you.

17:32Why this is bigger than one reconciliation

And look, this is bigger than one reconciliation. The finance pros who learn to direct the work don't just get their nights back. They become the ones their company can't afford to lose.