What Should You Do With the AI Prompts Your Team Keeps in Folders?

Turn the ones that matter into shared skills. A skill is a set of written instructions your AI assistant loads when a task comes up, with a clear purpose, a named owner and a version number. Test each change before it goes live, keep the current version in one place, and retire the copies. Your best prompts then become part of how the firm works, and you can say which instructions produced any piece of client work.

If your staff use an AI assistant every day, you probably know the situation. The prompt that drafts a good client letter lives in someone's documents folder. A better version sits in a Teams chat from last spring. Three colleagues have each tweaked their own copy, and nobody owns any of them.

What is a skill?

A saved prompt is a block of text someone pastes into a chat. A skill is the same knowledge written down properly: what the task is, when to use it, the steps, your house style, what it must never do, and any reference material it needs. The assistant picks it up when the task comes up, so nobody has to paste anything.

It also has an owner, a name and a version. A prompt in a folder has none of those.

Most business AI assistants now let you share instructions across a firm, under different names. Claude's Team and Enterprise plans, for example, let an administrator add a skill once for everyone, and Anthropic's help pages say users stay on the approved version until a new one is approved. Check what your own plan supports before you design around it.

What can a skill do?

More than most firms expect, and the simple ones are often worth the most. A skill used twenty times a week by ten people pays back quickly, even if it took an afternoon to write.

Simple skills, often an afternoon's work:

• write a client letter in your firm's tone, layout and sign-off
• turn rough meeting notes into your standard file note
• draft replies to the questions clients ask most, using wording you've already agreed
• check a draft against your style rules: spelling, required disclaimers, how you refer to products and fees

In the middle, a few days to build and test:

• read a client form and list what's missing before the meeting
• compare a document against a checklist, such as an engagement letter against your standard terms
• summarise a long file into a one-page review note, with each point linked to where it came from

More complex, built with care and checked by a person every time:

• read several documents, pull out the figures, apply rules that change every year such as rates and thresholds, and produce a structured pack with every number traced to its source
• prepare a first-draft report in a fixed format from a set of inputs, flagging anything unusual for review

The complex ones get the most attention, but most firms should start with the simple ones. They're quick to build, easy to test, and they show your people what good looks like.

What goes wrong when prompts live in folders?

The copies drift apart

Every time someone improves their copy, the firm has two versions. Within a few months there are five, and the best one belongs to whoever cares most. The quality of your output starts to depend on who you sit beside.

Nobody can say which version produced a piece of work

A client queries a letter, or a reviewer asks how a summary was produced. If the instructions behind it were a personal copy that's since been edited, nobody can say what the assistant was told. "We're not sure which version we used" is an uncomfortable answer to give a client, an auditor or the Central Bank.

A small change quietly makes results worse

Someone adds a sentence to make the output shorter. It works for the case they tried and breaks two they didn't. Nothing fails loudly. The drafts just get a little worse, and it can be weeks before anyone links the change to the drop in quality.

Client details end up in the prompt

People write prompts from real work, so real detail creeps in: a client's name in an example, a fee, a paragraph from a confidential file. The prompt is then copied, shared and pasted into tools nobody checked, and the client's details travel with it. GDPR's data minimisation principle (Article 5(1)(c)) says personal data should be limited to what's needed, and a prompt doesn't need it. Our post Is It Safe to Put Client Data into ChatGPT? covers which data belongs in which tool.

How do you turn scattered prompts into skills?

None of these steps needs special software.

1. Collect what's in use. Ask everyone to send the prompts they actually use.
2. Pick the ones worth keeping. Look for tasks that happen often, that reach a client, or where a poor result would cost you. Five good skills will do more than fifty half-kept ones.
3. Write each one up with a clear purpose and an owner. Merge the best parts of the competing copies. Give each skill a name and a short description that can't be confused with another, so the assistant picks the right one. The owner is usually someone senior in the area the skill serves.
4. Take out client details. Use placeholders or invented examples. If the skill will handle personal data when it runs, decide up front who can use it, how long its outputs are kept, and what your AI provider's terms say about the data.
5. Give it a version number. Every change gets a new number, the same everywhere the skill appears. Keep past versions, so you can always answer "what was the assistant using the day that letter was written?"
6. Test it before each change goes live. Keep ten or so real tasks with client details removed, and a note of what a good result looks like for each. Run each new version against them. A change goes live only if the results are at least as good. "It feels better" isn't the test.
7. Keep one copy and retire the rest. The current version lives in one place everyone uses, ideally loaded for the whole firm. Then ask people to delete their own copies, or the drift starts again within a month.
8. Review it on a schedule. Quarterly suits most firms. The owner asks whether the task, the rules or the regulation behind it have changed.

Is this a compliance matter or just tidiness?

Both. If an AI assistant drafts client communications, the instructions it follows are part of how your firm gives advice and service. Showing what those instructions were, who approved them and how they were tested is the same discipline you already apply to templates and sign-off.

It also helps with questions that are coming anyway. A short list of your skills, each with an owner, a version and the date it was last tested, is easy to hand over when a client, an auditor or your board asks how the firm uses AI.

Who should own the skills?

The people who own the work. The skill that drafts engagement letters belongs to whoever is responsible for engagement letters. Their job is to decide what good looks like, approve changes and keep the test cases up to date.

Someone also looks after the mechanics: the shared location, the version numbers, the archive and the testing. In a small firm that's often the operations lead.

What do we do when we build this for a client?

We collect what's already in use, help pick the skills worth building, write them up with owners and version numbers, and set up the tests each one has to pass before a change goes live. Skills that touch personal data follow the same data rules as everything else we build. We work the same way on our own skills. The full method is on our how we build page.

Where to start

Pick three tasks your team does every week and turn them into skills. Client work gets more consistent because everyone uses the best version, people stop hunting for the good prompt, and the firm can take on more with AI because each new use starts from instructions that are owned and tested.

If you want an honest picture of where your firm stands first, our free AI Readiness Self-Assessment takes ten minutes. A dploy.ai AI Operations Assessment maps where AI pays back for your business, including which of your team's prompts should become skills, for a fixed fee agreed before we start. Or book a short call and talk it through first.