How to Build an AI Prompt Library for a Marketing Agency

Written by Naveen Sehgal, Zylous Digital.

An AI prompt library for marketing is a shared set of tested instructions that a team reuses for recurring work such as keyword grouping, ad copy, search term reviews and monthly reporting. Without one, each person keeps their best instructions in a private chat history, and the quality of the output depends on who happens to be doing the task that week.

This guide explains what to collect first, how to write each entry so that someone else can use it, where to store the collection, how to test it and how to keep client information out of it. The advice draws on published guidance from Google, OpenAI and Anthropic.

Why one person's prompts do not scale

Most teams start informally. One person writes a good prompt for Google Ads headlines and reuses it from memory. A colleague writes a different version for the same task and gets a weaker result. After a few months several people are doing the same job in different ways, and nobody can say which version performs best or who changed what.

A shared collection also gives you a place to build in review. Google's guidance on generative AI content says AI can help with research and with structuring original work, and that a person should verify the facts and read through anything it produces before it goes live, down to titles, meta descriptions and structured data. Google also warns that using AI to produce many pages without adding value can breach its spam policy on scaled content abuse. A prompt entry that states what must be checked, and by whom, turns that review into a routine step.

Define what a good result looks like first

Anthropic's prompt engineering documentation advises setting clear success criteria, and having a way to test against them, before refining a prompt. In marketing terms, that means writing down what a good output looks like for each task. For a responsive search ad prompt, one criterion is checkable by anyone: Google Ads allows up to 15 headlines of 30 characters each and up to four descriptions of 90 characters each, so any output over those limits fails. For a search term review, the test might be whether the model flags the same wasteful queries that an experienced analyst found in the same report.

What to collect first

Begin with tasks your team repeats every month. Fifteen to twenty entries is a workable first version, because each one can then be tested properly. Group them by the stage of work they support:

Research: competitor analysis, audience research, keyword grouping and search intent classification.
Content: briefs, outlines, first drafts, meta titles and descriptions, and FAQ drafting.
Paid media: responsive search ad copy, negative keyword suggestions, search term reviews and Meta ad ideas.
Reporting: monthly summaries, explanations for sudden changes in performance, and recommended next steps.
Client communication: status updates, meeting recaps and proposal outlines.
Quality control: fact-check passes, brand voice checks, and repetition and readability checks.

How to write an entry that anyone can use

Google's Workspace guidance on prompting mentions four elements of an effective prompt: a persona, a task, context and a format. It adds that a prompt does not need all four, and that refining a weak result through follow-up prompts is normal.

OpenAI documentation likewise recommends explicit instructions and a clear structure that separates the role, the instructions, any examples and the background context. Those two ideas combine into a standard template for the whole team. Each entry should record:

Name and purpose: one line on what the prompt does and when to use it.
Inputs needed: what must be pasted in, such as a search term export, a client brief or a page URL.
Prompt text: written with a persona, task, context and format, and with square-bracket placeholders for details that change.
Success check: the test described in the previous section, so the reviewer knows what to look for.
Tool and version: the AI tool and model it was tested on.
Sample result: one real input and output with client details removed.
Owner and last tested date: who maintains it and when it was last run.

Two shorthand frameworks: CRAFT and ART

Prompt guides often compress the template above into an acronym, and CRAFT is the most common. Published versions agree on Context, Role, Action and Format, and differ on the last letter, which some guides read as Target audience and others as Tone. Add the audience or tone only when it changes the result, for example when the same analysis is written once for a client's marketing manager and once for their finance director.

ART is a shorter pattern for quick tasks. In this guide it stands for Act (the task you want performed), Request (the specific output you want back) and Terms (the limits, such as length or words to avoid). It leaves out the persona and the background, so it suits simple jobs such as rewriting one headline, and it needs extra detail added when the job is complex.

Neither is a formal standard, and published versions of CRAFT differ, so treat both as memory aids. The fields in the entry template above are what matter. A CRAFT or ART label can sit in the entry as a note on which structure the prompt follows. Whichever you choose, test each prompt on real inputs, because no structure guarantees a good result.

Worked example: a search term review prompt

The prompt below applies the persona, task, context and format structure to a common paid search job. Replace the items in square brackets each time it is used.

Persona: You are a senior Google Ads analyst. Task: Review the search term report pasted below and sort every term into one of three groups: keep, add as a negative keyword, or monitor. Context: The client is [CLIENT NAME], a [INDUSTRY] business. The goal is [LEADS OR SALES] at a target cost per conversion of [TARGET]. Do not recommend action on terms with fewer than [MINIMUM CLICKS] clicks. Format: A table with the columns term, cost, conversions, group and a one-line reason. State any assumptions you made after the table.

The placeholders keep client details out of the stored text, and the instruction to list assumptions gives the reviewer something specific to check.

Where to keep the collection

Use a tool the team already opens every day. A shared spreadsheet or a Notion database works, with one row per entry, one column per field above, and filters for workflow and tool. If you also build custom assistants, such as Custom GPTs, Claude Projects or Gemini Gems, treat the spreadsheet as the master copy and paste from it, so the two never drift apart.

Record the model name and version beside each entry. OpenAI advises pinning production applications to a specific model snapshot so that behaviour stays consistent, and building evaluations that measure prompt performance when model versions change. A marketing team is not deploying software, but the same reasoning applies: if the tool changes underneath a prompt, the entry needs to be tested again.

Test every entry, then keep testing

Run a new entry on at least three real inputs, for example reports from clients in different industries, and compare the results with work a person has already done. Note where it falls short. A prompt tuned on an e-commerce account can behave poorly on a professional services one, and that limit belongs in the entry. Once a quarter, run the most used entries again, update the test date and remove any that nobody uses.

Protect client information

Stored prompts should contain placeholders only: no client names, passwords, customer lists or unpublished financial figures. Check the client's contract before pasting real data into any AI tool, and check the tool's data terms for your plan. OpenAI states that by default it does not use inputs or outputs from ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu or its API platform to train its models. Anthropic says the same for its commercial products, including Claude for Work and its API, with an exception for conversations a user chooses to submit as feedback. Personal and free accounts are treated differently by both companies, and terms change, so read the current wording before relying on it.

Common mistakes

Copying prompts from the internet and adding them without testing them on your own work.
Leaving an entry without an owner, so nobody updates it when results drift.
Omitting the output format, which makes results hard to compare or paste into a report.
Sending AI-assisted work to a client, or publishing it, without a human fact-check.
Saving real client data inside stored prompts.

Starting this week

List the ten tasks your team repeats most often and write down what a good result looks like for each.
Write or collect one prompt for each task and test it on two real examples.
Put them in a shared spreadsheet using the fields described above.
Name an owner for each entry.
Book a 30-minute review for three months from now.

Where a shared prompt bank fits

A shared prompt bank is the first layer of a wider AI workflow. Zylous Digital applies AI across SEO, paid media and social media through its AI-first approach, and can help you set up a similar system for your own team.

Frequently asked questions

What is a shared prompt collection?

A shared set of tested instructions, organised by task, that a team reuses to get consistent results from AI tools.

How many entries should it hold at the start?

Fifteen to twenty is a workable first version, covering the tasks your team repeats most often. Add more once a task comes up repeatedly.

Where should the entries be stored?

In a tool your team already uses every day, such as a shared spreadsheet or Notion. Choose one with filters and version history so changes are traceable.