Your best prompts keep disappearing
Somebody on your team spends twenty minutes crafting a prompt that produces a genuinely excellent proposal draft. It works, the client is happy, and the prompt dies in that person's chat history. Three weeks later a colleague reinvents a worse version from scratch. Multiply that across every task your team uses AI for and you are paying the learning cost repeatedly while banking the benefit once.
A prompt library fixes this. It is a shared, versioned collection of your proven prompts, written as fill-in-the-blank templates, each with a named owner and a real example of good output. It turns individual wins into standard operating procedure, which is where the compounding value of AI actually lives for a small business.
Anatomy of a good library entry
A prompt pasted into a page is a start, but a reusable entry needs enough context that a new starter can use it correctly on day one. Give every entry the same eight fields:
- Name: following your naming convention (more on this below)
- Owner: the person accountable for keeping it working
- Tool and model: where it runs, e.g. Claude, ChatGPT or Copilot
- The template itself: with placeholders in curly braces, such as {client name} and {call notes}
- Required inputs: what the user must gather before running it
- A golden example: one real input and the output you consider good
- Version and last-reviewed date
- Known failure modes: where it goes wrong and how to spot it
The golden example is the field teams most often skip and the one that matters most. It sets the quality bar concretely, so nobody has to guess what good looks like.
Need a hand with this?
Our team delivers AI & Machine Learning for UK businesses — with a free initial consultation, transparent fixed quotes and no lock-in contracts. Tell us what you're working on →
Where to keep it
For most small teams a Notion database is the natural home: one entry per prompt, properties for category, owner, version and status, a gallery view for browsing and comments for feedback. If you live in the Atlassian world, a Confluence space does the same job; Microsoft-centred teams can use a SharePoint list or Loop workspace. Even a well-structured shared Google Doc beats the current alternative, which is nothing.
The AI platforms themselves offer packaging too: Claude Projects, ChatGPT's custom GPTs and Copilot agents all let you bake a proven prompt into a reusable assistant your team can open and use. Do that for your top handful of prompts, but keep the source text in your own library as well. If the prompt only exists inside a vendor's interface, you cannot audit it, migrate it or survive a plan change.
Naming conventions that scale
Names are retrieval. Six months in, a library of forty prompts called 'email one' and 'good summary prompt final v2 NEW' is a junk drawer. Adopt one pattern from day one: team, task, output, version.
- sales-followup-email-v3
- seo-meta-description-v2
- ops-meeting-minutes-summary-v1
- hr-job-ad-first-draft-v4
Three rules keep it clean: one prompt per job rather than one mega-prompt for everything; the version number lives in the name so nobody grabs a stale copy from an old message; and old versions are marked deprecated rather than deleted, so you can roll back when a rewrite underperforms.
Versioning and a review cadence
Prompts rot. Models are updated and respond differently, your services and prices change, and your brand voice evolves. Treat the library like any other SOP: every change bumps the version and gets a one-line changelog note saying what changed and why. The owner of each prompt is responsible for it still working, not the library as a whole.
- Book a 30-minute review monthly for the first quarter, then quarterly once stable
- Re-run each prompt's golden example against the current model and compare outputs
- Check outputs still match current prices, services and brand voice
- Retire prompts nobody has used since the last review
- Collect one new candidate prompt from each team member per session
That last item matters: the review is also your capture mechanism. Wins surface in the meeting instead of dying in chat histories.
Key Takeaway
Create a Notion database this week with eight fields per prompt: name, owner, tool, the template itself, required inputs, one golden example, version number and last-reviewed date. Name prompts by team, task and version (sales-followup-email-v3), deprecate rather than delete, and book a 30-minute monthly review to test each prompt against current models. Once a prompt is stable, write it into the relevant SOP so the workflow, not the individual, owns the quality.
From prompt to SOP
The endgame is not a tidy library; it is prompts embedded in documented workflows. Your sales SOP should say, at step three, 'run sales-followup-email-v3 with the call notes, then edit for tone before sending'. At that point quality stops depending on who is at the keyboard, onboarding a new starter gets faster, and the process owns the standard rather than the individual.
Start small this week: ask each person for their single best prompt, write the eight fields for those five entries, and hold your first review in a month. That modest start beats a grand taxonomy nobody fills in. If you want help designing the library or wiring prompts into your team's workflows and automations, our team can help.