Three People Have the Best AI Prompts and Nobody Else Can See Them
Photo by Photo by Justin Morgan on Unsplash on Unsplash
There's a quiet pattern in most UK organisations that have been using AI tools for more than a few months. The official rollout produced a training session, a policy PDF and a licence count. Then, underneath all that, individual practitioners built their own libraries — usually in a notes app, sometimes in a spreadsheet, occasionally in a Teams chat with themselves.
The prompts in those private notes are almost certainly better than anything in the official guidance. They've been refined against real work, they're tuned to the specific tone of a specific client, and they've survived the kind of trial and error that no training session replicates.
That's the good news. The bad news is that they're invisible, undated, unverified, and one resignation away from disappearing.
The Library Is Real, It Just Isn't Where You Think
Ask a team of ten whether they use AI to draft anything and you'll get a mix of enthusiasm and quiet non-answers. Ask the same team to share the prompts they actually use and you'll typically find three or four people doing the heavy lifting — the ones who've worked out that a well-structured instruction with a couple of examples in it saves twenty minutes on a Wednesday afternoon.
Few of them have shared it. Not because they're guarding an advantage, though occasionally that's part of it. Mostly because sharing a prompt means being asked to explain it, maintain it, defend it, and eventually be the person who gets the message in eight months asking why it "stopped working" after a model update.
The result is a shadow library: functional, fragmented and privately owned. It works right up until the person who owns it moves on, goes on parental leave, or simply stops caring.
Why Prompt Libraries Rot Faster Than Anything Else
Most knowledge bases die slowly. Prompt libraries die fast, for three specific reasons.
The ground shifts underneath them. Models get updated, defaults change, output styles drift. A prompt tuned in March can produce noticeably different output in September without anyone touching the file. There's no error message — the output just gets slightly flatter, slightly longer, slightly more generic, and people quietly stop using it.
Context is implied, not written down. The best prompts are full of assumptions: this is for our house style, this assumes the reader knows the product already, this assumes we're replying to a complaint rather than a query. Those assumptions are invisible to anyone else, so the prompt fails in someone else's hands and they conclude it doesn't work.
Nobody has been given the job. A shared prompt library with no owner isn't a resource — it's an accumulating pile. Within a quarter it has four versions of the same instruction, two of which contradict each other, and a new starter has no way to choose.
What a Light Routine Looks Like
You don't need a governance programme for this. You need one owner, one front door, one review date, and an explicit rule about what doesn't go in.
One owner, not a committee. A single named person who can say yes or no — usually someone with enough hands-on experience to judge whether a prompt actually works. This is a half-day-a-month job, not a role.
One front door. One place prompts are published. Not four. If your organisation has already standardised on a knowledge tool, use it — even if it's slightly awkward. A slightly awkward single door beats three elegant ones.
A date on everything. Every entry carries a "last checked" line and the name of the person who checked it. That's it. No version numbers, no changelog discipline, no approval workflow. A date and a name is enough to tell you whether to trust something.
A quarterly sweep. One hour, once a quarter, where the owner goes through the list and marks things as current, retired or needs a look. Anything that hasn't been used or checked in two quarters gets archived. Prompt libraries are a perishable good — archiving is a feature, not a failure.
Versioning for People Who Hate Versioning
If you try to introduce semantic versioning to a prompt library, you will lose the room in about ninety seconds. Practitioners didn't sign up to be release managers.
A simpler habit works better: keep the working version at the top of the entry, and below it, a short "what changed and why" note with a date. One line per change. If someone needs the old version, they can usually find it in the doc history — or, more realistically, they never will, and that's fine.
The point isn't archaeology. It's knowing which prompt to copy at 4:45pm on a Friday.
What Doesn't Go In
Be explicit about exclusions, because this is where light routines quietly become heavy ones. Client names, personal data and anything commercially sensitive don't belong in a shared library — full stop. In the UK, that's an ICO question rather than a workflow preference, and it's worth stating plainly rather than leaving it to judgment.
Beyond that, keep it practical: don't publish a prompt nobody has used on real work, and don't publish one you can't explain in a sentence. A library of thirty proven prompts beats a library of three hundred experiments every time.
Keeping Practitioners Onside
There's a version of this that fails, and it's the version where a central team announces a prompt repository, sets a standard, and asks everyone to migrate their private notes into it by Friday. That produces compliance in name and abandonment in practice.
Start instead with the three people who already have the best private notes. Ask them what they use, publish it with their name on it, and let the quality do the persuading. Give them the credit rather than the maintenance burden — at least at first.
The goal isn't tidiness. It's making sure the twenty minutes saved on a Wednesday afternoon is saved by everyone on the team, not just the person who happened to figure it out first.