Prompt library vs training: practise on your own work
A prompt library gives you other people's tasks, and most training gives you exercises you will never repeat. Pick one real task from your own week, time how long it takes now, then redo it with AI and compare. That comparison is the only thing that tells you whether anything changed.
A prompt pack sitting in a shared folder does not change how the work gets done.
People ask whether to buy a prompt library or book training. Both can leave you exactly where you started, because neither one makes you redo a task you already do.
What follows is the method we use in prompting sessions with bank and insurer teams. Pick one real task. Get a before number. Redo it with AI. Compare. It takes about a week of normal work, no offsite needed.
What happens when you actually open a prompt pack
Most packs have a long list of prompts in them. Forty are for marketing copy you do not write. Twenty are for coding. The rest are about email.
You open it before a call, scroll for something close to the memo you are writing, find nothing, and write the memo by hand. The prompts were written for somebody else's work. Hunting for a match takes longer than just doing the task.
The same thing happens after training staff were required to attend. Everyone practices on the trainer's example, which is clean and finishes inside the exercise. Then Monday arrives with a messy version of a task nobody demonstrated, and the file stays closed.
Prompt library vs training: the same failure point
Both run into the same wall. Neither one is attached to a task you repeat.
A library only helps when you already know what you are looking up. Training only helps when there is a real task waiting on Monday, and usually nobody has named one.
Whichever you buy, the outcome comes down to one thing. Someone takes a task they already do, redoes it with AI, and pays attention to what happens.
In our sessions the practice work is always the participant's own material. Their own claims letters, their own onboarding checklist, their own weekly report to a boss. The result is different because the work is real and the same task comes round again next week.
Pick the first real task you will use AI on
Pick the one you do most often, not the one that matters most.
Good candidates are things you do at least weekly, that produce text or a bit of analysis, and that you are allowed to get wrong once. Drafting a summary of branch staff feedback. Turning meeting notes into a decision log. First-pass quality checks on a document that someone senior reviews anyway.
Bad candidates are the ones people usually pick first: a board paper, a regulatory response, anything where the review chain will absorb the time you saved and more.
Write the task down in one line. "Summarise the weekly complaints list into five themes with examples." That line is your test case. Judge every prompt against it and nothing else.
A prompt library is useful at this point. Once you have a task you repeat, you can look up how someone else phrased it and tell quickly whether it fits.
Get a before number, even a rough one
Most teams have no baseline. The question "did it work" comes back as a shrug, and a shrug does not survive a conversation with your boss.
Before anyone opens an AI tool, do the task once the old way and write down two things: how long it took, and how many times it came back for correction. That is enough. A rough number you actually wrote down is worth more than an exact one you never collected.
We ran a series of five-day design sprints on customer-facing web journeys for a large insurer. Journeys that used to take six months or more were designed, tested with customers, and live in four weeks. Completion rates went up 80 percent and the old drop-off points disappeared. Those numbers exist only because the client measured the old journey first. Without that before number, there would be nothing to prove the sprints worked.
Same logic for prompting. Twenty minutes down to six on a weekly task is a small honest number, and small honest numbers hold up.
Check the Monday after, then fix the task
A week later, ask the plain question. Did anyone open the AI tool for that task, and did it help?
If the answer is no, the task was wrong. Usually it came round too rarely, or it needed sign-off before anyone could touch it. Pick another task and run it again.
If the answer is yes but the output needed heavy editing, look at the task itself before blaming the prompt. Often the input is the problem. Notes nobody structured, a form with free-text fields that branch staff fill in differently every time, a turnaround time that is long because the work sits in someone's inbox, not because the writing is slow.
If the customer journey is broken, AI will just make the broken part faster. Fixing the input usually saves more time than any prompt will.
What to do this week
- Write down one task you do at least weekly, in one sentence.
- Do it the old way once and note the time taken and the number of corrections.
- Redo the same task with AI, keeping the prompt that worked in a file you will actually reopen.
- Compare the two numbers and write the difference down, even if it is small.
- Next Monday, check whether you used it again without being reminded.
More on AI prompting
Questions people ask
Is a prompt library useless then?
No. A prompt library is a decent reference once you already have a task you repeat. It just will not change how you work on its own, because the prompts are written for someone else's work.
What should I measure if I have no baseline?
Time one round of the task by hand before anyone touches AI, and note how many times it came back for correction. A rough number written down beats a precise number you never collected.
How do I pick which task to practise on?
Pick something you do at least weekly, that produces text or analysis, and that you are allowed to get wrong once. Frequency matters more than importance.
On-Off Group trains teams, tests products with real customers, finds where a transformation has stalled and builds what gets it moving, for banks, insurers and enterprises in the Philippines, since 2015. How we help with ai prompting.
