FlowsBook a callSign inCreate free accountSign up
AI employees and automation

A prompt library for teams: tested AI instructions with versions and an owner

Keep the AI instructions that work in one place, use them in every flow, see who changed what and when, and test the flows that depend on one before its next run.

Part of the guide AI employees

By the xMap team · · 5 min read

A prompt library for teams is one shared place for the AI instructions your business relies on, such as how you score a lead or how you write to a customer, with a version history, so every flow reads the same words and a change made once reaches all of them. In xMap Flows it lives under Settings, Instructions. Each entry has a name, a line on what it is for and its words; AI blocks pick it from a list, and every save keeps the version before it.

Without one, the same idea gets pasted into ten flows and quietly drifts: one copy was sharpened after a bad week, another still has last year's target market, a third was shortened by someone in a hurry. This article explains how a shared library fixes that, what belongs in it, how to change an instruction safely, and a made-up example from a recruitment firm. The guide to AI employees covers the wider setup.

How a saved instruction is used

Three AI blocks take saved instructions: the prompt of xMap LLM, the goal of an AI agent, and the question of Jev decision. Under the block's own instructions box, Use saved instructions lists the library; pick up to five.

  • The saved words come first, then the block's own words. Keep the block's box for what is particular to that one flow: this city, this kind of place, this output field.
  • You can also place an instruction anywhere in a block's settings with a tag such as {{instructions.how_we_score_a_lead}}; the instruction's page shows its tag.
  • Each run records which instructions it used and which version, so an odd answer can be traced back to the words in force that day.

What belongs in the library

Put in the instructions several flows share and that carry a judgement your team has agreed on. Leave out the details that change from flow to flow.

Good library entryWhy it is sharedWhat stays in each flow
How we score a leadSales, partnerships and outreach flows all rank companiesThe kind of company and the city
How we write to customersEvery message an AI drafts should sound like one firmThe reason for this particular message
What counts as a good siteExpansion and property flows judge the same wayThe format of the store being planned
What we never sayPrices, promises and competitor names stay out everywhereNothing; this one is the same for all

A handful of entries usually covers most of a team's AI blocks.

A good entry reads like a page from your sales playbook: who you serve, what makes a strong fit, what rules a company out, and an example or two. Short, specific and written as instructions, not as a wish list.

Changing an instruction safely

The point of a shared instruction is that improving it once improves every flow that uses it. That is also the risk, so the library builds in a routine:

  1. Open the instruction and edit its words. Saving makes a new version with your name, the time and a short note on what you changed.
  2. Press Compare with now on any earlier version to see exactly what moved.
  3. Press Test the flows that use it. It lists them, shows what a test run of each costs, and runs them. Nothing is sent and nothing is remembered.
  4. Read the results. If one flow now behaves worse, either refine the wording or keep that block on the earlier version with Keep this version on its chip until you have time to look.
  5. If the change was a mistake, Restore makes an older version the newest again. Nothing is lost.

Flows pick up the new words from their next run. A block set to keep a version stays on it until someone presses Use latest.

Who may change what

  • Members add and edit instructions; viewers can read them.
  • An organization admin can Lock an instruction, after which only admins change it. Lock the ones that carry policy, such as what an AI may never promise.
  • An instruction that any flow uses cannot be deleted. Its page shows the flows that use it, so you can take it out of them first.
  • If a flow names an instruction that has gone, it runs on its own words and its step says so, rather than failing silently.

A made-up example

A recruitment firm in Johannesburg runs three AI employees that score companies: one finds companies posting jobs for finance roles, one watches new offices opening in business districts, and one drafts first messages to hiring managers. Each had its own idea of a good client pasted into its prompt.

The team writes one entry, What makes a good client for us: companies with 50 staff or more, several open roles at once, roles the firm places often, and no in-house recruitment team mentioned in the posting. All three flows pick it from the list and keep only their own detail in their boxes.

Two months later the firm stops placing junior roles. The lead consultant edits the one entry, notes no junior roles, and tests the three flows. The message-drafting flow now writes oddly short notes, so she keeps that one block on the previous version for a day, rewrites one line, and switches it back to the latest. One change, three flows aligned, and a history that shows why.

Moving instructions between teams

A flow saved as a file carries the instructions it uses. Import it into another organization and its instructions join that library, or the receiving team's own entry of the same name is used instead. Ask xMap also knows the names and descriptions of your entries, so when it builds a flow from a sentence it can pick the right one rather than inventing new wording.

Questions

Which blocks can use a saved instruction?

The xMap LLM prompt, the AI agent's goal and the Jev decision question. Any block's settings can also include an instruction through its tag.

What happens to flows when someone edits an instruction?

They use the new words from their next run, unless a block was set to keep a particular version. Testing the flows that use it before the next run shows any surprises.

Can I undo a bad change?

Yes. Every save is a version; Restore makes an older one the newest again, and nothing is deleted.

Can anyone change our most important instructions?

Members can edit by default, but an admin can lock an entry so only admins change it.

Templates that do this

Read next

More on ai employees and automation · Every article · Templates

Put it to work

An AI employee in xMap Flows does this on live data, on its own timetable. Start from a template or describe it in plain words.