prompting

AI is sloppy until it isn't

By Samer Azar, Fractional CFO · 2026-07-03 · 6 min read

Key Takeaway: A non-human-sounding AI is nothing but a poorly instructed one. You can fix today's output by hand, or you can fix the instructions once and never fight it again. The second one is slower, and it is the whole game. Do it for long enough and the machine stops being sloppy right when you least expect it.

Dear reader,

A little while ago I sent out a report my AI had mostly written. I was proud of it. It was accurate, it was thorough, it was fast.

The people who read it told me it sounded like a machine wrote it.

They were right. It was correct and it was cold. Every sentence did its job and none of them sounded like a person. So they asked me to humanize it, and I sat with what that word actually means.

Two ways to fix it

There were two roads in front of me.

The first was to open the document and rewrite it line by line in my own voice. That works. It also solves nothing. Next month I would be doing it again, on the next report, from scratch.

The second was slower. Instead of fixing the words, fix the instructions. Teach the machine how I actually sound, once, so every future draft comes out closer to right.

Being the productivity nerd I am, I did both. The line-edit for the report going out that week. The instructions for every report after it.

What I actually gave it

Here is the part worth stealing. The "humanize it" instruction was not vague. I handed the AI the same rules I write by, in plain words:

• Write to one person, not to an audience. The single reader living this right now.

• Read it back and ask if it sounds like a person talking or a brand publishing. If it is the brand, cut it.

• No corporate words. No leverage, no synergies, no paradigm shifts.

• Say the true thing in a short sentence and let the reader draw the conclusion.

• No hedging. Delete every "I think" and "I believe."

That is the whole trick, and it had nothing to do with a smarter model. Last month I blind-tested the models I already pay for against my own documents, and the expensive one lost. The model was never the decision. I told the one I had who it was writing for and how I talk, and it stopped sounding like a robot because it was finally being instructed by a human.

Sloppy until it isn't

This is the part nobody tells you about building with AI.

For weeks, sometimes months, it feels like a bad trade. You tweak an instruction and get a slightly less wrong answer. You connect one more source. You fine-tune the same prompt for the fifth time. It is thankless and slow, and there is no single moment where it obviously clicks.

And then one day it clicks.

A question I would normally have spent two hours on came back close to answered. It reasoned cleanly and showed its sources so I could check every line. I still went through it carefully, and it held.

Long-time readers know this song. I have written before about the file that cures AI amnesia, and about the private archive that quietly compounds in the background. A few of you wrote back asking what is actually in mine. So here it is.

Before that question ever reached the AI, a step ran first and pulled seven things into one brief:

• rules.md: the standing instructions I wrote months ago about how we think and what a good answer looks like

• company.md: the shape of the business, its entities, its fiscal year, the people in it

• this week's live numbers: pulled straight from the accounting system into a fresh table

• decisions.md: a running log where every real call gets written down with its date and its reason

• one outside benchmark: for the single ratio that actually mattered

• glossary.md: so it defines margin and cash the way we define them, not the way the internet does

• the question itself: framed with the right scope instead of dropped in raw

Those files sit in one folder, plain markdown, read top to bottom before the model says a word. Nothing fancy, just a desk laid out the same way every time. The model did the last ten percent. The ninety percent was that folder.

Half of what is in there lives for a moment, and half of it lasts. This week's numbers and the question in front of it are gone the second the window closes. The rules, the history, the decisions, the glossary all stay, because I wrote them down for it to keep. Feed it only the first half and you get a genius with amnesia, brilliant and brand new every morning. The good answer needed both. It knew who this company was, and it knew what had just changed.

The sentence that made the whole thing land came out of decisions.md. A note from eight months earlier about why we had changed a supplier arrangement, one line, the kind no human in the room would have remembered was there. The machine remembered, because I had written it down for it to keep. That is what a long memory buys you. It never forgets the useful thing you forgot.

That is the moment the grind pays for itself, all at once, the way it always seems to happen.

Where you have seen this before

If that pattern sounds familiar, it should. It is the same shape as cash.

Internal cash flow feels like nothing for a long time. One small lever, then another, week after week, and none of it looks like much. Then it compounds, and the business is suddenly funding its own growth. Small and invisible, until it isn't.

The machine works the same way. The reps look wasted right up until the day they clearly weren't. The only real mistake is quitting during the boring part.

A well-instructed AI, fed your own company's context, that surfaces the few moves that actually matter instead of a wall of numbers. That is the whole idea behind Alex the CFO, and it is exactly this craft pointed at your cash. I am opening it to a founding group of twenty.

Don't be a knocker-upper

Before alarm clocks, people paid a knocker-upper to walk the street and tap on their window to wake them. A real job, until a one-time purchase did it better and it vanished.

Editing your AI's voice by hand, every time, is knocker-upper work. You can do it forever, or you can instruct it once and let it wake itself up. Don't be a knocker-upper.

Take care of your cash, and it takes care of almost everything else.

Samer

P.S. I want to steal from you now. Reply and tell me the one instruction that changed how your AI works for you. I read every reply, and the best ones end up teaching the rest of us in a future issue.

Know someone fighting to make AI sound human and actually useful? Forward them this. It might save them a month of guessing.

Not subscribed yet? Join the CFO Lab for the weekly build.

Want to share it and get rewarded? Use your personal referral link: https://blog.alexthecfo.com/subscribe?ref=PLACEHOLDER