This week I ran an on-site training with a client team, teaching Claude live and building examples in front of them. One of those examples broke in a way I didn't expect. I built the same board deck twice, same prompt, same source data, and got two decks that were almost identical but not quite.
That's not a bug. It's how Claude is built. But it raised a question worth answering properly: which of my own Claude workflows drift like that without me noticing, and what do I actually do about it?
This week's free section covers why it happens, using that deck and a reconciliation task that drifted the same way. The paid section gives you the framework for locking down the tasks that need to come out the same every time, plus a prompt you can use to build one yourself.
Updates and useful links
Claude in Action, Cohort 2 is underway; we're two sessions in. If you want a seat in the next cohort, the waitlist is here.
Want Claude training for your team? I run corporate sessions built around how your team actually works. Reply to this email, and we'll set up a call.
New this month: a few clients have asked me to just build the thing instead of teaching them to build it. So I'm opening up two options: build-with-you and built-for-you; more details coming soon.
If you've got a workflow you want automated, a skill you want built, or you know exactly what Claude should be doing and just don't have the time to do it yourself, reach out. I'll tell you straight if it's something I can help with.
Claude Outputs: Same but Different
I built the deck for Meridian Advisory Group, the fictional client I use in examples. Same prompt, same numbers, run twice. The two decks weren't identical. The commentary on the slides was worded differently, and the layout wasn't quite the same either. Both were fine on their own. Neither was wrong. But if I'd sent one version to a board and the other to an audit committee, someone would have asked why they didn't match.
I've seen the same thing on the numbers side. A client and I built a monthly expense reconciliation together. Two months in a row, Claude completed the task correctly both times. But the spreadsheet came out shaped differently each month, different column order, different subtotal placement. The reconciliation itself held up. The format didn't.
Here's the part that's easy to miss. This isn't Claude making a mistake. It's Claude doing exactly what it's built to do.
Anthropic draws a useful line in its own engineering guidance on building agents. A workflow is a system where the model follows a predefined path. An agent is a system where the model decides its own path as it goes. Agents are genuinely better for open ended work, a first draft, a scenario you haven't thought through yet, a deck you're still shaping. The same flexibility that makes agents good at that work makes them the wrong tool for anything that has to come out the same way every time.
A quick myth to clear up. Telling Claude to be consistent, or turning the randomness down, does not fix this. The output can still change between runs even when the settings stay the same. The fix isn't a setting. It's locking the method itself.
Which means deciding, task by task, whether you want Claude to think fresh or repeat itself. If you're exploring a problem, let it think fresh, that's the whole value. If you're producing the same accrual, the same reconciliation, the same board slide every month, don't ask Claude to redo the thinking. Build the method once, check it, and reuse it.
Here's what that looks like across four kinds of finance work.
Formulas. Don't ask for the calculation fresh each time. Ask Claude to build the formula set once, check it, then just drop new numbers in. “Calculate this month's accrual” becomes “build the accrual formulas once, I'll reuse them every month.”
Templates. Don't describe the look each time. Lock a template and only feed it new data. “Build me a board deck, make it blue, add a variance chart” becomes “use this exact template every month, only the numbers change.”
Scripts. For calculations too complex for a spreadsheet, ask for a script once, verify it, then run that same script instead of asking Claude to redo the math.
Matching rules. This one isn't about format; it's about judgment. Instead of asking Claude to match transactions to invoices and trusting it’s read each time, write the rule once: invoice number first, then amount within a cent, then date within three days. Claude applies the same rule every time instead of deciding fresh, and you can see exactly why anything didn't match.
For finance work specifically, this matters more than in most fields. A marketing team can live with a slightly different deck every time. A board can't, not when two versions of the same numbers are sitting in two different inboxes.
And when an auditor asks how a number was produced, here is the formula, here is the script, or here is the matching rule, and here is when we last checked it, that's a five minute conversation. Claude figured it out is not an answer you can stand behind, even when the number was right. Locking the method isn't just about getting the same output twice. It's what you hand an auditor.
None of this is about trusting Claude less. It's about being precise on which tasks need a fresh mind and which need a fixed one. Get that distinction right once, and you stop reconciling the same drift every month.
Knowing the difference is one thing.
Building the method is another.
Below: the exact framework for locking a script or a spreadsheet, two full worked examples you can copy today, and the one detail that determines whether any of this actually runs.
Closing Thoughts
An on-site training this week reminded me why I still enjoy this part of the work. Watching someone go from “I don't trust what this gives me” to “I know exactly why it gave me this” is really rewarding.
If you build a locked script from the prompt above, I'd like to hear what task you picked, and whether it caught anything you didn't expect.
We Want Your Feedback!
This newsletter is for you, and we want to make it as valuable as possible. Please reply to this email with your questions, comments, or topics you'd like to see covered in future issues. Your input shapes our content!
Want to dive deeper into balanced AI adoption for your finance team? Or do you want to hire an AI-powered CFO? Book a consultation!
Did you find this newsletter helpful? Forward it to a colleague who might benefit!
Until next Tuesday, keep balancing!
Anna Tiomina
AI-Powered CFO