How to Build a Prompt Library for Your Profession (That You'll Actually Use)

Disclosure: this post was written with AI assistance.

Most people who use AI assistants at work end up with the same problem: a handful of prompts that worked really well once, scattered across chat histories, never found again. A prompt library fixes that. It's not a list of clever tricks. It's a small set of reusable instructions tied to the tasks you actually do every week.

Here's how to build one that stays useful instead of turning into another forgotten document.

Start from your task list, not from prompts

Open your calendar and your sent folder for the last two weeks. Write down every recurring task that involves reading, writing, summarizing or deciding. For a project manager that might be status updates, meeting notes, risk logs and stakeholder emails. For a recruiter: job ads, screening notes, rejection emails, interview question sets.

Pick the five tasks that are both frequent and tedious. Those are your first library entries. Ignore everything else for now; a library with five prompts you use daily is worth more than one with two hundred you never open.

Use a fixed structure for every prompt

Prompts that work reliably tend to share the same parts. Give every entry these fields:

• Role and context: who the model should act as and what situation it's in. "You are helping a hospital ward manager write a weekly staffing update for senior nurses."
• Input: what you will paste in, clearly delimited. "Notes are between triple quotes."
• Task: one clear instruction. "Write a 150-word update."
• Constraints: tone, length, format, things to avoid. "Plain language, no acronyms, bullet list for open issues."
• Output format: exactly how the answer should look, ideally with a short example.

Writing prompts this way makes them easy to adjust. When output is too long, you change one constraint instead of rewriting everything.

Add placeholders, not hard-coded details

Replace the parts that change with obvious placeholders: [CLIENT NAME], [DEADLINE], [PASTE NOTES HERE]. This turns a one-off prompt into a template. Keep placeholder names consistent across the library so you learn them once.

A good test: could a colleague in the same role use the prompt without asking you anything? If not, there's context missing.

Store it where you already work

The best storage is wherever you'll see it. Options that work well:

• A text expander (Espanso, TextExpander, or your OS's built-in replacements) so typing ;status inserts the whole prompt.
• A notes app with one note per prompt and tags by task type.
• A shared doc or wiki page if your team will use it, with a table of contents at the top.

Avoid storing prompts only inside the AI tool's history. Histories get long, and some tools limit how far back you can search.

Keep a short log of what works

Next to each prompt, keep two lines: the date you last changed it, and a note on what went wrong before the change. "Added 'do not invent figures' after it made up a budget number." This is the most valuable part of the library over time, because it captures failure modes specific to your work.

Build in verification steps

For anything factual, add an instruction that makes checking easier: "List any claims that I should verify, separately at the end," or "Quote the sentence from my notes that supports each point." AI tools still produce confident errors. A library that bakes in verification protects you from the worst of them, and keeps you honest about what you hand to clients or managers.

Never paste confidential client data, personal data or anything covered by an NDA into a tool your organisation hasn't approved. Put that rule at the top of the library.

Examples by profession

To make this concrete, here are starter entries for a few roles:

• Accountant: "Turn these bank transaction descriptions into categories from this list: [CATEGORIES]. Output a table with the original text, category and a confidence of high/medium/low. Mark anything ambiguous as 'review'."
• Teacher: "Create five multiple-choice questions on [TOPIC] for [AGE] students. One correct answer, three plausible wrong answers based on common misconceptions. Include the misconception each wrong answer tests."
• Sales rep: "Here are my call notes. Write a follow-up email under 120 words that restates the customer's problem in their words, confirms the next step and date, and asks one question."
• Developer: "Review this function for edge cases only. Don't rewrite it. List inputs that would break it and explain why in one line each."

Notice that each one is narrow. Narrow prompts give predictable results.

Review monthly, prune ruthlessly

Once a month, look at which prompts you actually used. Delete or archive anything untouched for two months. Merge near-duplicates. Promote any ad-hoc prompt you typed more than twice into the library.

Common mistakes

• Collecting instead of using. Downloading huge prompt lists feels productive but rarely changes how you work.
• Vague roles. "You are an expert" adds little. Specific context ("for a small dental clinic's patient newsletter") adds a lot.
• No output format. If you don't specify the shape of the answer, you'll spend time reformatting.
• Never testing. Run each new prompt on three real examples before trusting it.

Summary

A prompt library is a working tool: five to twenty structured, tested templates for your recurring tasks, stored where you'll see them, with notes on what broke and how you fixed it. Start small this week with your most tedious task, and grow it only when a prompt has earned its place.

---

If you want a starting structure instead of a blank page, I put together a small, inexpensive, AI-assisted kit on this topic: AI prompt libraries for specific professions. Everything above works without it. Other kits: all products.