AI Prompt Templates Library

Why a template beats a blank box

An empty prompt field is a bad way to start. Not because writing prompts is hard, but because you describe what you want in the order it occurs to you, which is rarely the order that produces a good answer. You lead with the topic, remember the audience halfway through, forget the format entirely.

A template fixes the order before you type. It has slots for role, task, context, constraints and output shape, so the only work left is filling them in — easier than inventing structure and content at once.

It is also a record of what worked. When a colleague gets a usable result, the template is how it gets passed on: not as advice, as something you can copy and run. And it exposes gaps — if a marketing template has a [target audience] slot you cannot fill, the gap is in your brief.

Adapting one properly

The most common failure is running a template unchanged. Every bracketed placeholder is a question you are being asked. Leave one in and the model either invents an answer or writes around it, and you get something generic. Replace all of them, then add the thing no template can contain: your specific situation.

A template as it comes:

And the same template actually filled in:

Same skeleton, completely different output. The context block is where the value is — the failed old guide, the tickets, the first task. None of that ships inside a generic template.

One check before you run it: could a competitor paste this prompt and get something useful? If yes, add more of your own situation. For why these blocks work in this order, see how to write AI prompts.

Building a library of your own

The library here is a starting set. Your own is the one that gets used, because it is full of prompts tuned to your product and your recurring tasks.

Organise by task, not by tool. "Rewrite a support reply" is findable; "Claude prompts" is not, because next month you will be using something else. A structure that survives contact with reality:

• By job to be done — drafting, editing, summarising, analysing, image generation.
• By recurring deliverable — weekly report, release notes, ad variations, meeting recap.
• A "scratch" pile for prompts that worked once and have not earned a permanent slot.

Give each entry three things: one line on when to use it, the prompt with placeholders marked, and a note on what it is bad at. That last one saves the most time — "good for first drafts, always invents statistics" tells a colleague how much to trust the output.

Keep image prompts separate. They follow different rules — subject, style, lighting, composition rather than role and constraints — so mixing them in makes both harder to search. Build those with the image prompt builder and keep a parallel set for Midjourney.

Versioning what works

Prompts drift. Someone tweaks a line, the output gets worse, nobody remembers what the original said. Treat a working prompt the way you would treat working code.

The lightweight version: keep the old prompt under a "v1" heading before you edit, date it, add one line on what changed and why — "added the no-exclamation-marks rule, output was too breathless". Fifteen seconds, and you can roll back.

Change one thing at a time. Rewrite the role, tighten the constraints and switch the format in one pass and you learn nothing from the result. Adding worked examples is a high-leverage single change; few-shot versus zero-shot prompting covers when it pays off.

Re-test your best prompts when you switch models. A rule that stops one model rambling can produce clipped answers elsewhere.

When a template is the wrong tool

Skip the library when the task is genuinely one-off, when you are exploring rather than producing, or when the request is short enough that scaffolding outweighs content. "Summarise this in three bullets" needs no role and no constraints block; overstructuring a small ask often produces a stiffer answer than the plain question.

Templates also fail when the real problem is that you do not know what you want yet. Have the conversation first, then turn the outcome into one.

Do that conversion deliberately. When a one-off produces something good, go back to the prompt and ask which parts were about *this* request and which apply to the next of the same kind. Swap the specific parts for placeholders, keep the structure fixed, note the situation it is for. That is a template built from evidence, not guesswork. The prompt optimizer will tighten it before you save, and anything reused often enough is a candidate for a system prompt.