How to build a prompt library your whole team can reuse: fields, owner, review rule and retirement

A prompt is the written instruction you give an AI tool to make an image, a video or a voice. A prompt library is the shared place where a team keeps the instructions that work, so the next person does not start from a blank box or pay for the same trial and error again.

Most guides to this show a list of columns and stop. This page gives you the columns, a filled-in example, the rules that keep the library trustworthy (an owner, a review step and a retirement date) and the extra fields that image and video work needs. It is tool-neutral. For where to store the finished images and videos themselves, see where to keep AI images, videos and prompts, which has a short prompt sheet. To file new outputs there without copying by hand, see moving AI output into your asset library automatically. This page goes further on teams and rules.

Fields: copy this table

Published guides on team prompt libraries, from AI Operative Supply and TextExpander (vendor pages, so read them as vendor advice), agree on a core set: name, use case, owner, approved tool, inputs, output format, an example, last updated and status. TextExpander (a tool for saving text snippets you can insert anywhere) adds variables, tools tested and last reviewed. I would take that core and add the image and video fields below, which those guides do not list.

Field	What goes in it	Why it matters
ID and name	P-014, "Product on white, soft shadow"	Lets people refer to it without pasting it
Use case	One sentence: when to use it, and when not	Stops misuse
Owner	One named person	Someone is responsible for keeping it good
Tool, model and version	For example "Image tool X, model Y, version Z"	The same words give different results on different models
Prompt text with [variables]	The words, with blanks in square brackets, such as [product] or [colour]	Makes it reusable
Guard line (what must not change)	"Keep the label, shape and colour exactly as in the reference. No text added."	Prevents the usual failures
Reference image or seed	File name or link; the seed number if the tool shows one	Lets you repeat a result
Aspect ratio and length	9:16, 5 seconds, or 4:5 still	Prompts often depend on the shape
Example result	Link to the best output file	Shows what "good" means
Last test: date, tester, result	2026-11-04, Sam, 4 of 5 passed	Shows the prompt still works
Cost per run	Credits or money per try, from the screen at the time	Teams then see the real cost per approved result
Rights and consent flag	Yes or no: does it involve a real person's face or voice, a brand or a client product?	Flags what needs permission
Client allowed	Which clients accept AI work and which do not	Avoids a nasty surprise
Status	Draft, Approved, Retired	The one filter everyone uses
Last reviewed and next review	Dates	Keeps the library from going stale

If you want a minimal version, my priority order is: name, use case, model and version, last tested date, owner, status. That is my judgment, not a standard.

One worked entry

This example is made up to show the level of detail. It is not a recommended prompt for any particular tool.

Field	Entry
ID, name	P-014, Product on white with soft shadow
Use case	Catalogue-style shot of a boxed or bottled product. Not for products with small print on the front.
Owner	Priya
Tool, model, version	Image tool X, model Y, version 2 (check the name on screen when you run it)
Prompt	"[Product] standing on a plain white surface, soft light from the left, gentle shadow, clean background, centred, [angle] view."
Guard line	"Keep the product shape, colour and label exactly as in the reference image. Do not add text, props or other objects."
Reference	client_product_front.png
Aspect ratio	4:5 still
Example result	Link to the approved file in the shared drive
Last test	2026-11-04, Priya, 4 of 5 outputs passed (one had a changed label)
Cost per run	As shown on screen at the time; record it each test
Rights flag	No people. Client product used with the client's permission.
Client allowed	Clients A and C yes; Client B no (asked for no AI images)
Status	Approved
Next review	2027-02-04

Notice the guard line and the "not for" note in the use case. Those two are what stop a new team member from using the prompt where it fails.

Where to keep it (a sheet first)

Start with a shared spreadsheet (Google Sheets or similar). It is searchable, shareable, free or nearly so, and you can add columns as you learn. Move to something else only when the sheet stops working. Comparing the options:

Option	What it is	Pick it when
Shared spreadsheet	Rows and columns in a cloud file	A small team, as a rule of thumb of mine, not a tested limit
Notion or Airtable	Notion is a notes and database workspace. Airtable is a spreadsheet-style database with forms and views	You want filters by client or tool, forms for submitting a new prompt and an approval view. AI Operative Supply names Notion and Airtable for this. A prompt manager such as Portkey (a service whose docs describe folders, permissions and version tracking) is a further step up
TextExpander or a snippet tool	Saves text you insert by typing a short code	The prompts are text only and used many times a day. It does not hold images or test results
A folder of documents	One file per prompt	Rarely. Search and review get hard

Whatever you choose, give edit access to the owners and approvers only, and view access to everyone else. Version history matters, because someone will overwrite a good prompt. Link each prompt to its best output in your asset library, so the library shows results and not just words.

Two things should never go in a shared prompt. First, client confidential data such as unreleased product details, customer names or private numbers. Many AI services use what you submit in ways you may not expect: Higgsfield's Terms of Use, for example, say your content, inputs and outputs may be used to train and improve its models, and that different terms apply to customers under an Enterprise Agreement. Read does your AI tool train on client work before you paste client material anywhere. Second, details of real people. Describe a role ("a barista in a green apron") and keep the consent record and photos in a separate, access-controlled place.

Review rule and retirement

A library without rules fills up with prompts nobody trusts. Four rules keep it clean:

1. Test before approve. A prompt is Draft until 3 to 5 outputs have been generated and saved, and the results are scored against a short checklist (shape right, label intact, no warped faces or hands, no stray text). Pass at least 4 of 5 for an approved image prompt. For video, use the same idea on clips. The pass mark is my suggestion, not a standard. Raise it for ads. The quality control checklist gives you a list to score against.
2. One named approver. A different person from the owner signs off, and the date goes in the sheet. This catches blind spots.
3. Review on a schedule. The vendor guides suggest monthly or quarterly. I would use quarterly for approved prompts and re-test two or three of them each time, plus after any model change.
4. Retire by default. TextExpander suggests archiving after 90 days. I would apply that to prompts nobody has used in 90 days. Mark them Retired, do not delete, and move them to an Archive tab with the reason.

Before a new client project, run through the client-allowed column and the rights flag. If a client's approval is part of the process, see client sign-off on AI drafts.

Model changes: why prompts go stale

This is the part most prompt library advice skips, and it matters more for images and video than for text. The tools change their models often, rename them or shut them down. A prompt that gave clean results last quarter can give worse ones now, or stop being available.

A clear example: OpenAI's help page says its Sora video service ended for the web and app on 26 April 2026 and for the API on 24 September 2026. Any library that listed Sora as the "approved tool" had to be updated. I cover that in Higgsfield vs Sora. Platforms that host many models also add and remove them, so the same applies if you use a platform such as Higgsfield, an online AI image and video service, rather than one vendor's own tool. The model names on screen are the ones to record.

What to do:

• Store the model name, the version and the date for every prompt, as the table above does.
• After any switch, re-test the top 10 prompts and update their last test date and result.
• Mark prompts "model-specific" or "works across models". Test the second kind on two models before you believe it.
• Keep the working version of a prompt, not only the idea. A prompt for one tool is rarely a clean copy-paste into another. For starter wording on a particular platform, see Higgsfield prompts and templates, and for a shared look across tools see keeping a brand look across AI images and video.

Start in one afternoon

1. Create the sheet with the minimal columns: ID, name, use case, owner, tool and model, prompt, guard line, example result link, last test, status.
2. Ask each person for their three best prompts. Put them all in as Draft.
3. Pick the five most-used ones and test each with 3 to 5 outputs. Approve those that pass.
4. Write the rules (test, approver, quarterly review, 90-day retirement) in the first tab, so people see them.
5. Add the extra fields (cost, consent, client allowed) the first time a client project needs them.

It is better to have 12 trusted prompts than 120 unsure ones.

Who should skip this

• You are one person who makes a few images a month. A single note with your best prompts and the model name is enough.
• Your team uses AI mainly for one-off, creative exploration. A library of fixed prompts can slow that down. Keep a notes page for ideas instead.
• Your prompts are all very short text-only tasks. A snippet tool may serve you better than a full library.
• You have no one to own it. An unowned library goes stale in a few months. Do not start it until a person agrees to be the owner.

How I'd approach this

I would start with a plain shared sheet, six columns, and the four rules above. I would test before approving, name an owner and an approver, re-test after every model change and retire what nobody uses. Then I would add the image and video fields only when a real project needs them. I would keep client secrets and real people's details out of it from day one.

If your team outgrows the sheet and wants a proper shared library with approvals and links to the finished files, Blur Studio can build the library and approval step around the tools you already use.