How to version prompts for an enterprise assistant
A prompt version for an enterprise assistant states which system message and which fixed prompt are in effect, who approved them, and how to return to the previous version. Document versioning says which file is the current text. Wrong-answer intake says what happens after someone marks a reply wrong. This note is the fixed wording the assistant applies each time it answers.
On this page
- Keep prompt versions apart from document versions and from wrong answers
- A system message is a set of instructions, and it does not guarantee the model will follow them
- Version the fixed text, not the question a colleague just asked
- A new version takes effect when the named person has approved it
- Returning to the previous version is written by the same approver
- If a wrong answer points at the prompt, name the version that was live
- Conflicting instructions, or a missing fallback, stay in draft
- The prompt is being managed when the version, the approver, and the previous path are written
Keep prompt versions apart from document versions and from wrong answers
Document versioning says which file on the store is current, and who may mark the previous draft withdrawn. Wrong-answer intake says that after a reply is marked wrong, one line holds the question, the wrong reply, and the correct source, and the next step is to change the document, recut the passage, retest retrieval, or change the prompt the reply relied on. Prompt versions answer a different question: which text of the system message, and of the fixed prompt wrapped around a colleague's question, is in effect, who approved it, and which version a return points back to. The three can point at the same answer. See How to manage document versions in a knowledge base and How to record and fix a wrong knowledge-base answer.
The path, version, and section a reply must carry are in How to attach sources to a knowledge-base reply. Who may open which set of documents is in How to set access levels on a knowledge base. The delivery scope sits under Internal knowledge base and assistant. A prompt version does not rewrite the current-file mark, and it does not open a second door for wrong answers.
A system message is a set of instructions, and it does not guarantee the model will follow them
Microsoft Learn on system messages for Azure OpenAI, read on 2 November 2026, shows no publication date. The page says it is about the system message, sometimes called a system prompt or metaprompt, for chat experiences on Azure OpenAI chat models. It defines a system message as instructions and context given to the model to guide responses, typically to define the assistant's role and boundaries, the tone, the output format, and the safety and quality constraints for the scenario. It says a system message may be one short sentence or several lines of rules. The rhyme example and the JSON example on the page are examples. They are not the text of this assistant.
The same page says a system message influences the model and does not guarantee that every rule is followed. It still has to be tested and revised, and layered with other mitigations such as filtering and evaluation. A chat API sends messages with roles such as system, user, and assistant. The system message usually comes first and is the highest-level instruction for that conversation. The page says the instructions should be unambiguous, the rules should not conflict, and the fallback should be explicit when information is missing or the request is out of scope. It also says responses can vary across models and versions, and that behaviour can change when user content conflicts with the system instructions, especially in a long conversation. That is the page's account of Azure OpenAI chat models. It is not an instruction to buy that service.
Version the fixed text, not the question a colleague just asked
The question a colleague sends this time is the content of the user role for that turn. It changes every time, and it is not given a version number. What is versioned are two fixed texts: the system message, and the fixed prompt the company wrote to be sent with the question. If there is no fixed prompt, the cover says "no fixed prompt". It is not left blank. Each text has a version, the date it became current, and a path where it is kept.
An instruction pasted into chat does not change the current version until it is written into one of those two texts. A draft sits beside the current text, marked as a draft, and the assistant does not apply the draft. The current file in the library is still marked under the document-versioning note. The prompt version only records which document-set version it was tried against. It does not replace that current-file mark.
A new version takes effect when the named person has approved it
The person who may approve a new version is named on the cover. The page does not name that person. Without the line, a draft cannot become current. In one change, the approver marks the new text current, marks the previous version withdrawn, writes the date, and points at the path of the new version. The withdrawn text stays readable, so a reply from that time can be compared with it.
Before approval, the record says what was looked at: which rows of the wrong-answer log, or which question set the retrieval note already prepared. If that scope is blank, the approval is not finished. How a wrong answer is logged, and how retrieval is retested, stay in those notes. See How to record and fix a wrong knowledge-base answer and How to evaluate knowledge-base retrieval. This page only records which batch of records the version number matches. It does not calculate a new score.
Returning to the previous version is written by the same approver
A return marks the withdrawn previous version current again, and marks the version now in effect as withdrawn. Both happen in one change, signed by the same approver, with the date and the reason. The reason points at a wrong-answer row, a retrieval record, or a conflict that has been written down. A blank reason means there is no return.
The return does not rewrite replies already sent, and it does not delete the withdrawn text. A sent reply that needs a correction goes through the wrong-answer note. The prompt only decides which version the next answer uses. Saying in chat to use the old sentence, without this signature, leaves the current version unchanged.
If a wrong answer points at the prompt, name the version that was live
If the next step on a wrong-answer line is to change the prompt, that line names the prompt version that was in effect, and the version that replaces it after approval. Without the version that was live, the cause is not written as the prompt. It is written that the version was not yet recorded. Recutting, changing the document, or retesting retrieval is still chosen in the wrong-answer note. It is not re-decided here.
The system message and the fixed prompt are two versions. The wrong-answer line says which of them changed. If both changed, it records two version numbers. If only one changed, the other version number stays as it was, and the same change says it was not changed.
Conflicting instructions, or a missing fallback, stay in draft
The page lists conflicting instructions among the common pitfalls, for example asking to be brief and to be comprehensive with no order of priority. It also lists an overly long system message, because a longer message uses the context and leaves less room for the user's content. On the draft, the company marks sentences that conflict, or marks that there is no fallback when information is missing or the request is out of scope. Until those sentences are revised, the approver does not mark the draft current.
The page says a system message does not guarantee that every rule is followed, and that responses can change with the model version. If a model name is kept, it is written on the cover with the date it was read or put into use. The model version is not used as the prompt version. The prompt version refers only to the text. If the model changes and the text does not, a line still records the date of the model change and the approver. Otherwise the current prompt is not treated as approved for the new model.
The prompt is being managed when the version, the approver, and the previous path are written
If the system message has no version, if the approver is unnamed, or if the withdrawn previous version has no path, the assistant does not apply that text. If the cover neither says there is no fixed prompt nor gives it a version, the fixed prompt is not applied either. The colleague's question is still accepted. The reply says the prompt version has not been written.
Three checks show the prompts are being managed: the current system message and the current fixed prompt each have a version and a date; a new version and a return are both written by that approver in one change; and a wrong answer that points at the prompt matches the version that was live. A blank item means the next answer does not apply text that has not been approved.