Skip to content
Interface language: English
Audit / Signal

A company prompt library: how do you know which instruction is in use today?

Which instruction does your company's AI use today? Organize prompt versions, private copies, owners, tests and restoration of an earlier way of working.

Article language: English

Aeneia illustration: A woman holds a sheet showing a leaf shape beside an open drawer decorated with an iridescent pattern.
In this article

The demonstration produced a good result. Two days later an employee got a different answer from the same tool, although the question looked similar. The team checks the model, documents and settings. Eventually it finds that the demonstration used an instruction stored in a private note, while the application still loaded an earlier version. The company has a prompt library. The problem is that everyone keeps it somewhere else.

A prompt is an instruction or set of messages directing a model's work. In a company it becomes part of a process when a task's result depends on its contents. People should know what it is for, who may change it and which version was actually used. A folder called “best prompts” answers none of those questions, although it can grow very impressively.

One task needs a recognizable instruction

Imagine a fictional studio using AI to prepare a working summary of a customer enquiry. The tool should extract the goal, materials provided and unresolved questions. It should not price the work or promise a date. The library's first resource might be an instruction for precisely this task, named for the result it produces.

Add a short description of the input and output. Does it receive the entire message or a cleaned form? Is the result a note for a person or data consumed by another application? These differences affect use. Copied text without a description may look fine while directing the model to do different work.

Separate permanent instructions from material belonging to a particular case. A customer's name, enquiry and current catalog should not accidentally be saved in the active instruction as an example for every later request. When reviewing a change, check whether the author improved a rule or pasted a one-off situation into it. Such a mistake can persist for a long time.

Keep the dependencies too: response format, tools, sources and the way the result is handed on. An instruction may require a document that the particular application version cannot access. If a dependency involves an MCP server, check the tools it offers and their descriptions for changes, using the article on identifying changes to MCP tools (in Polish) as a guide, so the library does not describe a different integration from the one in use. Identical prompt text alone does not establish that the whole process is identical. Reproducing a test requires the conditions under which it was used.

A company AI-use policy can set general rules for responsibility and data, distinct from a task-specific instruction. The article on an organisation-wide AI-use policy (in Polish) discusses the policy in more detail. The library should apply them to a specific resource: this instruction, this task and this input scope. Employees then do not have to reinterpret the entire policy whenever they run an ordinary form.

The saved version and the version in use are different facts

A version number identifies immutable content. A “production” label says which version was chosen for current work. Separate these concepts. If the production label moves from version five to version six, an earlier execution should still point to version five. Otherwise history starts changing meaning with every correction.

LangSmith prompt-management documentation describes versions, Staging and Production environments, and labels pointing to particular commits, or saved states. The environment pointer can change without changing the historical state itself. This distinction helps check which instruction the application fetched at a given time.

Do not assume the application always fetches the resource again. It may cache a copy, load it at startup or use a file included at deployment. Check the actual route. A change in a management panel does not prove that the next user request receives the new content.

In the fictional studio, label a test with the instruction version and application version. If a result needs explanation, the team should be able to reconstruct the instruction actually used, including dynamic parts. A link to the document currently open is not enough. It may have changed since the task ran.

You also need a simple rule for private copies. People can use them for experiments in an agreed environment, but the result should be recognizable as an experiment. If a copy becomes the basis of daily work, it must go through the established assessment and approval process. Otherwise the official library will describe only what the company no longer uses.

A variable name is not enough to reconstruct a result

The studio's instruction may contain a place for the current list of services. Check who fills it and from which source. Two tests with the same fixed prompt text but a different service scope can produce different correct results. If only the instruction number was recorded, that difference cannot be explained. Reproduction also needs the state of the relevant material.

How the message is assembled is part of the check too. An empty field is not always equivalent to saying material was not provided. Default text may be received as customer content. Test an example with no attachment, an incomplete description and several services. Establish which absences the application should recognize and which the assistant should name in its output.

History does not require copying all data unconditionally into a shared library. The instruction, source-version identifier and actual customer message have different roles and may need different access. Decide what is needed to explain a result, where it may be stored and who can read it. Examples available to the whole team can use fictional material.

Check exports as well. A file sent for assessment may contain a confidential example even if the fixed prompt itself is harmless. The approver should know exactly what is leaving its existing storage location. Calling it an “instruction library” does not change the attachment's contents or the recipient's permissions.

Aeneia conceptual photography

The storage tool has a useful life too

Choose a library mechanism after checking the provider's current documentation. The OpenAI deprecation schedule records the deprecation of reusable prompt objects announced on 3 June 2026 and the planned shutdown of v1/prompts on 30 November 2026. The schedule was checked on 29 October 2026. Check the dates again before implementation: an announced shutdown states the provider's plan, not proof that it was later carried out.

The current OpenAI prompting guide recommends storing instructions for new work in version-controlled application code and subjecting changes to review and tests. That approach has different organizational requirements from a shared document edited directly by every employee. Establish who prepares the wording and who introduces it into the application in use.

This does not mean a person who understands customer service must change code personally. They can prepare a proposal and cases showing the problem. The person responsible for the application moves the change into the correct resource. Approval should concern the content that will actually be used, not a similar version from a conversation or presentation.

A smaller team can use a simpler workflow if it can distinguish states. A proposal, a checked version and the version in use should not be mixed in one field without history. Maintaining three clear stages may be simpler than later discovering who corrected a sentence appearing in five different notes.

Before choosing an elaborate tool, follow one real change from a reported problem to a user performing the task. See whether the history is understandable and whether an earlier version can be restored. Only then assess what the current way of working lacks.

Approve a change based on behavior

Suppose the studio wants a shorter working note. The author removes several sentences from the instruction and gets a concise result on one enquiry. They still need to check that unresolved questions are retained. The shorter text may improve readability while removing the condition that kept the model from inventing a date.

Attach an example of the problem and the expected effect to the proposal. “Shorten it to the goal, materials and questions” identifies a recognizable change. “Optimize the prompt” leaves too much open to interpretation. The reviewer needs to know what should look different and which behaviors must remain after the change.

Run AI response regression tests on known important cases and prepare a test for the new problem. Compare results against the stated criterion. A record saying “tests passed” should lead to the material version and observations, not a memory of a successful conversation last week. See the article on AI response regression tests (in Polish).

Identify who can withdraw the change. If notes begin losing customer questions, the team needs to restore a checked state and mark cases requiring another review. Returning to the old prompt does not automatically correct results already produced. Their subsequent handling needs a separate decision.

Keep the reason for the decision in a project decision log or the associated change history. For guidance on that record, consult the project decision register guide (in Polish). The next person should understand why an apparently redundant sentence was added. It may guard against an error invisible in the shortest example. Without that knowledge, the same safeguard can be trimmed away every month.

The owner is responsible for scope, not every comma

The instruction owner should understand the task, decide on proposals and know where the resource is used. They need not write every change personally. They should establish who can propose changes, who checks effects and who approves use. This lets the tool improve without granting everyone access to every production version.

Treat an instruction found in a public library as material to assess. Check whether it requests data the task does not need and assumes access to tools absent from the company. Copying it does not approve its scope or effects. Your own examples should show whether the proposal solves the identified problem.

Permission to edit a draft can be separated from permission to introduce it into daily use. Rehearse that boundary on an ordinary change, without an outage: who sees the proposal, who can approve it, and what does someone denied access see? If the mechanism is unclear on a quiet day, an urgent correction is unlikely to make it orderly.

When the author leaves or responsibility changes, hand over the test history, dependencies and limits too. The instruction file is only a small part of the knowledge. The new person needs to know when the prompt works correctly, where work was deliberately left to a person and how to report a problem. Otherwise they inherit a resource's name without a way to maintain it.

The library should also show retired instructions. They need not be available for accidental execution, but their history helps explain earlier results. Mark what replaced them and whether private copies remain. Telling people “use the new version” will not itself move text saved in a shortcut on someone's computer.

If considering fine-tuning a model on company examples, retain the instruction version associated with the training material. See also the article on fine-tuning (in Polish). If the process uses routing between AI models, check every path actually used. How a task reaches a particular model is discussed in the article on routing work between AI models (in Polish). The same text may behave differently with another set of tools and data.

Start a conversation with Aeneia Audit with one result and a question: which instruction produced it? If answering requires searching chat, a drive and the memories of three people, we have already found work to organize. Recover that trail first. The next “brilliant prompt” can wait a little.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear