Every small-business ChatGPT guide hands you a list of fifteen prompts. Almost none of them answers the question that decides whether those prompts are still working in March: where does a prompt live after you close the chat?
Here is the uncomfortable version. The prompt is not the asset. The asset is the prompt plus the notes about when it works, what it produces, and which version is current. A prompt pasted from a chat thread has none of that. It is a photocopied key without a door.
The full prompt library with all fifteen copy-ready prompts sits in the complete guide on tanit365.com. What follows is the part guides skip: how to run prompts as procedures, so the library survives staff changes, expired subscriptions, and your own memory.
Every prompt gets one page
One prompt in one chat thread is a slot machine. One prompt on one page is an asset.
A prompt page is deliberately boring: the prompt text at the top, then three short sections underneath - when to use it, what a good output looks like, and the inputs you must fill in. If a teammate opens the page cold, they should be able to run the prompt without asking you anything. That is the whole test. A page that needs a narrator is not a procedure; it is a souvenir.
Keep the page named after the job, not the tool: weekly-content-calendar, not chatgpt-prompt-number-three. Tools get swapped; jobs persist.
Name outputs the way you will search them
A prompt that writes a good product description is worthless in month four if nobody can find the eleven descriptions it already wrote. Outputs need names before they need folders: business name, document type, date, version. That one line means an intern can find every product description from October without asking. Search is the real interface of a prompt library - not the chat history, which hides everything behind one login.
Add an evaluation gate before the prompt ships
A prompt is not done when the first output looks good. It is done when the output survives a check.
Before a prompt becomes team procedure, run it three times on real inputs and grade each output against a short list: facts correct, format matches the template, tone matches the brand, no invented numbers. If any output fails a check, the fix goes into the prompt text, not into your hopes for next time. That single habit is the difference between a prompt that quietly degrades and one a new employee can run on day one.
Version the prompt like anything else that matters
When a prompt changes - and it will - the old version does not get deleted. It gets dated. Version history on a prompt page costs one extra line per change and answers the question that saves hours in month six: did the outputs get worse because the model changed, or because the prompt changed? Without a dated line, that question is unanswerable, and teams respond by re-prompting from scratch, which erases the accumulated fixes that made the prompt good in the first place.
Own the vault, not the subscription
The quiet failure mode of every prompt library is ownership collapse. The prompts live in one person's chat threads; the outputs live in a Downloads folder; the shared document lives in one Google account. When that person is out sick, the library is out sick.
The fix is boring: the prompt pages and their outputs live in a folder the business owns, synced outside the machine that created them, with version history. Whether that is a synced drive or a cloud folder with previous versions, the rule is the same - no single login, and no single laptop, is load-bearing. Cloud storage priced once instead of rented monthly fits a prompt library well, because the library's value compounds while its size barely grows; the storage guide on tanit365.com compares the plans that actually work that way.
The thirty-minute audit
Run this once: list every prompt your team actually used last month. For each one, answer four questions - does it have a page, is the output named so it can be found, did anyone grade an output this month, and does the current version live somewhere that survives a lost login? Most teams score zero on all four for most prompts. That score is not a criticism; it is the roadmap. Thirty minutes moves the library from chat threads to operating procedure, and the next model upgrade becomes a non-event instead of a reset.
FAQ
Where should prompt libraries live for a small business?
In a folder the business owns, synced outside the machine that created it, with version history. Chat threads are a workshop, not a vault: a lapsed subscription or a departed employee can erase the library. If you also keep the outputs there, previous versions protect a prompt when someone overwrites it.
How many prompts should a small team maintain?
Fewer than feels impressive. Five prompts with pages, names, and an evaluation gate outperform thirty loose prompts in chat threads, because the team can actually find, run, and improve the five. Start with the workflows that repeat weekly.
Should prompts be shared with the whole team or kept private?
Shared - as pages, not as chat invites. A prompt locked in one person's thread dies with that login. A prompt page with when-to-use notes and filled-in examples is usable by a teammate on day one without a training session.
What is the fastest way to tell a good prompt from a bad one?
Run it three times on real inputs. A good prompt produces usable output at least twice with nothing changed; a bad one needs a different lucky phrasing each time. Grade against a fixed short list - facts, format, tone, no invented numbers - so the comparison is honest.
Do prompt libraries need paid tools?
No. A shared document plus a synced folder covers most teams for months. What matters is the discipline - one page per prompt, named outputs, a grading step - not the software. Paid storage only earns its cost when the library and outputs need to survive outside one machine.
