The Prompt Hoarder

An accounting ops person I know found something last week while auditing their AI implementation. They’d deployed a system to handle “client engagement letters” — the foundational document that outlines the scope and fee for a client engagement.

She asked the team “which version of the engagement letter prompt are we using?”

Audit director said “I’ve got the one I like in my notes.”

Tax manager said “mine’s in the shared folder under Templates.”

An associate said “I’ve been using the original one from two months ago that someone sent me.”

Three different versions. Three different people. All in active use. And when a compliance issue showed up in one, it was fixed differently than it had been fixed in another two months prior. Ops person was literally fixing the same mistake in three different places.

Yeah. Nothing improved. Nothing learned. Everything fixed twice.

What Examples and Guardrails actually are

Quick frame first. We wrote a piece called Stop Hiring AI. Start Building It. that introduced the Professional Recipe — seven things any operator with sense gives a new hire on day one. Two of those seven are Examples (Ingredient #4) and Guardrails (Ingredient #3).

Here’s what they do.

Guardrails are the boundaries — never quote a price. Never reply to a service complaint without me. Always include the compliance disclaimer. They keep the AI from producing the wrong kind of output.

Examples are reference outputs — “here are three engagement letters that landed. Here’s what made them work. This is the standard.” They show the AI what good looks like.

Here’s the thing — together, they form the spec. Spec is how the AI knows what it’s supposed to produce. When the spec is scattered, the AI never converges on a standard. Different people running the same tool produce different outputs because the boundaries aren’t shared.

Why prompts scatter

Here’s what happens. Team gets an AI tool deployed. Engagement letter template, let’s say. Spec comes with guardrails — compliance language, fee structure, scope clarity.

Audit director uses it for a week. Works. But the director thinks “this is fine, but for audit clients I’d prefer slightly different tone. Let me adjust it.” They edit the prompt locally. Now they have their version.

Tax manager uses the original. But they notice the compliance language doesn’t fit tax engagements. They edit it. Now they have their version.

An associate gets onboarded by the tax manager. Tax manager gives them the adjusted version. Not the original. Associate doesn’t know an original existed. They now have their version.

All three think their version is official. All three are “improving” the prompt. None of them are feeding the improvement back to the system.

Right? This isn’t laziness. It’s rational. Specialist knows their client base. Their version is probably better for their work. The problem is systemic, not personal.

The failure mode — the prompt never improves

Here’s what the ops person experiences. Same problem shows up in two different engagement letters, written by the tool six weeks apart.

She fixes it in the current version. She emails the team. “Hey, engagement letter prompts — we had this compliance issue. It’s fixed now.”

Three weeks later, audit director gets flagged for the same issue. Director fixes it in their local version.

Three weeks after that, the associate gets the same flag. They fix it differently — they didn’t know there was already a fix.

Ops person is firefighting the same fire three times.

Why? Because there’s no system. There’s no single source of truth. There’s no maintenance cycle. There’s no feedback loop that says “this prompt is the official version, improvements go here, everyone runs from the same spec.”

There’s just scattered copies. Each person tending their own garden. Nothing syncing back.

The compliance cost in professional services

Here’s where this gets sharp — in an accounting firm, this matters more than in other verticals.

Engagement letters are brand assets. First thing a client reads about your firm. Tone, scope clarity, fee structure — they set expectations. If your firm’s engagement letter is inconsistent — one version too casual, one version too formal — clients notice. Firm doesn’t look professional. Firm looks scattered.

Worse — compliance. An engagement letter is a legally binding document. Has to include specific language. Can’t vary by person. But when every specialist has their own version, that variance is inevitable. Ops person has to audit the field regularly, catch inconsistencies, and patch them.

Ops Firefighter is maintaining three systems instead of one.

How to know if your prompts are hoarded

Simple test. Ask your team “what’s the current version of our [key document] prompt?”

Three different answers? Hoarded.

Three different people pointing you to three different folders? Hoarded.

Someone says “I have it in my notes”? Hoarded.

Someone says “I don’t know, I’ve been using the one [person] gave me”? Hoarded.

Nobody can tell you who owns it or where the official version lives? Hoarded.

How hoarding happens

Before the Monday Move, here’s how prompts scatter even when you don’t intend them to.

Audit director gets the official prompt. Uses it for two weeks. Thinks “this is fine, but for my clients I’d prefer warmer tone. I’ll adjust it for my book of business.”

They adjust it. Locally. No announcement. They’re not trying to break the system — they’re just being good at their job. Their version is genuinely better for their specific client base.

Tax manager gets the original. Doesn’t fit tax engagements. They adjust it. Also locally.

An associate gets onboarded by the tax manager. Gets given the tax manager’s adjusted version. Doesn’t know an original exists.

Now you have three people each maintaining a system they think is official. They’re all doing good work. Problem is systemic. Does that make sense?

This is why centralization isn’t mean. It’s the only way the system gets better.

The Monday Move

So. Find every version of one critical prompt your team uses. Engagement letters. Client communication templates. Billing documentation. Pick the high-value one.

You’ll probably find three to seven versions.

Read them. Pick the best one. That’s your official version. If there isn’t a clear winner, combine the best elements from all versions into a new official version.

Assign ownership — one person owns that prompt. They’re responsible for it. They’re the person who accepts improvements and updates it. Make this explicit. Tell them “you’re the keeper of this prompt. All improvements come through you.”

Tell the team — this is the official version. This is where we all work from. When you find something that needs fixing, you tell the owner. You don’t fix it locally. If you want to adjust it for your specific practice area, you propose that to the owner first.

Redirect everyone to the official version. Destroy the old copies or archive them clearly as “old — do not use.” This is important. As long as the old versions exist, people will use them.

Next month, measure — how many versions are in active use now? My bet — one.

The shift

Ops Firefighter’s job is to keep systems clean. Hoarded prompts aren’t a prompt problem. They’re a system design problem.

You have five specialists, each maintaining their own version of a shared asset. That’s not specialization. That’s fragmentation.

So. Own your prompts. Integrate them. Maintain them. When improvements come in, they go into the system. Everyone runs from the same spec. The ops work doesn’t multiply.

That’s the shift.

Framework: The Professional Recipe — Guardrails (Ingredient #3) and Examples (Ingredient #4). Related failure modes: The Prompt Hoarder — Examples and Guardrails scattered instead of integrated.

Companion piece: Stop Hiring AI. Start Building It. — the parent framework. This one closes the gap on Ingredients #3 and #4 specifically.