---
title: "How to build a brand prompt library for your marketing team | Rainbrand"
description: "A brand prompt library is a shared set of tested prompts for the assets your team makes every week. Keep the brand out of each prompt and have it fetch current brand context instead, put an acceptance check at the end of every prompt, and test each one on a normal and an awkward task. With four prompts we tested, including what went wrong."
url: "https://www.rainbrand.com/blog/build-a-brand-prompt-library/"
---

By Siddharth Borderwala

Published 17 Nov 2026

11 min read

# How to build a brand prompt library for your marketing team

A brand prompt library is a shared set of tested prompts for the assets your team makes every week. Keep the brand out of each prompt and have it fetch current brand context instead, put an acceptance check at the end of every prompt, and test each one on a normal and an awkward task. With four prompts we tested, including what went wrong.

## TL;DR

A brand prompt library is a shared set of tested prompts for the assets your team makes most often. Write each prompt to describe the task and fetch the brand’s current guidance and files, rather than pasting a copy of the guide into it. End every prompt with the check its output must pass, test it on a normal task and an awkward one, and write down what went wrong.

## What a brand prompt library is

Prompt libraries are turning up in marketing job ads. Block wants a [B2B Content Marketing Lead](https://block.xyz/careers/jobs/5422288008?gh_jid=5422288008) with “experience building AI-assisted content systems, prompt libraries, research workflows, repurposing engines, or content operations processes.” Viral Nation’s [Director, Content Marketing](https://job-boards.greenhouse.io/viralnation/jobs/5281628007) will “Build reusable prompts, voice guides, templates, and agentic workflows so research, drafting, repurposing, and QA run as a repeatable engine rather than one-off efforts.” Lovable’s [Brand Designer, AI & Ops](https://jobs.ashbyhq.com/lovable/df79d83b-e366-4998-afdd-694d06183d75) will keep the visual system as “style references, fine-tunes/LoRAs, prompt libraries, not just a static guidelines deck”.

A prompt library is a shared set of prompts for recurring tasks, each tested and kept up to date, so nobody starts from a blank chat. Generic ones exist: Jasper’s public [Prompt Library](https://www.jasper.ai/prompts) sorts prompts by use case and marketing channel. A brand prompt library adds the part a generic one can’t: every asset it produces has to be on-brand. That depends less on how a prompt is worded than on where it gets the brand from, and on whether anyone has checked what it produces.

## Keep the brand out of the prompt

The common way to make a prompt on-brand is to paste the brand into it: the hex codes, the voice rules, [“never say revolutionary”](https://www.rainbrand.com/blog/brand-messaging-framework/). It works for a week. Then the palette changes, or a product is renamed, and every prompt in the library carries the old version. Forty prompts are forty copies of the guide to update, and the people using them can’t tell which are current.

Write prompts that describe the task and fetch the brand. The prompt says what to make, for whom, in which format and from which facts. The brand’s guidance and files come from one maintained source at the moment the prompt runs. If your agents are connected to Rainbrand, which gives AI agents a brand’s guidance and files, that means asking for a [Brief](https://docs.rainbrand.com/guides/request-brief): the guidance and files that apply to the task. If not, point the prompt at one shared, maintained file or folder rather than pasting from it.

- Instead of Use our colours #1A1A1A and #FF5A00 and our friendly, witty voice.

    Request a Brief for \[brand\] for this task, and use its guidance and files.

- Instead of Mention that setup is 40% faster.

    Take every claim from \[source\], and list where each one came from.

- Instead of Make a social image with our logo and the headline.

    Generate the picture without text or a logo. Place the logo file the Brief chose, and set the headline in the Brief’s font.

- Instead of Make sure it’s on-brand.

    Before showing me the result, check it against \[the specific test\], and tell me what you changed.

Fetching also corrects the prompt. On Rainbrand’s [Cursor setup page](https://www.rainbrand.com/integrations/cursor/), a prompt asked for a launch section for Nothing’s Headphone (1) Pro with the product name “in Nothing’s dot-matrix type”. The Brief that came back said the brand had no dot-matrix font file, only the dot-matrix NOTHING wordmark, and that the wordmark mustn’t be used to letter the product name. It also marked what the brand didn’t settle, such as the price and feature callouts, and said to take those from the product page.

![Cursor showing the prompt that requested a Rainbrand Brief for Nothing, and the Brief's colours by role, copy guidance and not-settled notes](https://www.rainbrand.com/integrations/cursor/brief-colour-copy.webp)

The prompt at the top, and part of the Brief it fetched, in Cursor on October 31, 2026. The “Not settled” lines are what the brand leaves to the task.

A prompt with the brand pasted in can only repeat what its author believed. Fetching also works across brands: an agency can keep one library and name the brand as a variable, because the brand isn’t written into the prompt.

## Structure each entry

Organise entries by the job they do, not by tool: launch email, feature post, webinar promotion, alt text, sales one-pager. Group the jobs by channel, so a social lead finds their prompts in one place, and keep one entry per job rather than five near-duplicates from different people.

Each entry needs more than the prompt text. Demand Spring’s [guide](https://www.demandspring.com/blog/marketing-ai-prompt-library/) lists use case, variables, examples, model settings, owner, version, evals and risk notes. For a brand library, record these:

| Field | What to write |
| --- | --- |
| Job it does | One line, such as “announces one feature on LinkedIn” |
| Use it for, and not for | Where it works, and where it would invent things |
| Inputs | The brackets someone fills in |
| Prompt | The text, ending with its acceptance check |
| Brand context | What it fetches, such as a Brief for the task |
| Facts | The sources it may use, and nothing else |
| Tests | A normal task and an awkward one, with brand, model and date |
| Known failures | What went wrong in testing, in a sentence each |
| Owner and version | A named person, a number and the date |

The acceptance check belongs inside the prompt, not only in the entry around it. People copy the prompt, not the entry, so a check that lives in a separate field never runs. Each of the prompts below ends with “Before showing it to me, check…”, and in every test that line did real work.

Keep the library somewhere with an edit history and a named owner: a shared doc, a Notion database or a Git repository. The format matters less than being able to see who changed a prompt, when and why.

## Four prompts, tested

We tested four prompts on November 16, 2026, each in a fresh Claude session connected to Rainbrand, on real brands and public sources. Each got a normal task and an awkward one: a misspelled product name, a claim the source doesn’t make, Japanese text, or copy that doesn’t match. Tested here means we ran each one and recorded what happened, including what went wrong; it doesn’t mean they’re safe to run without a person reading the output. [Download all four](https://www.rainbrand.com/downloads/blog/build-a-brand-prompt-library/brand-prompt-library-starter.md) in the entry format above, with the tests and failures filled in.

### Launch email

Launch email

```
Request a Rainbrand Brief for [brand] for a [length] email introducing [product] to [audience]. Show me the parts of the Brief you’ll follow. Write the email, claiming only features that the Brief or [product page URL] supports. Before showing it to me, check it: list each claim beside the words on the source that support it, remove any claim you can’t source and tell me you removed it, and confirm every product name is written exactly as the brand writes it.
```

We ran it for Teenage Engineering: a 120-word email introducing OP–XY to people who already own an OP-1 field, citing the [OP–XY page](https://teenage.engineering/products/op-xy). The email came back lowercase, as the Brief asked, with every feature claim matched to words on the page. The check earned its place: the agent removed a line from its own draft, “sync it all with your OP–1 field over the multi-out jack”, because neither the page nor the Brief says it. It also wrote both names with the en dash the brand uses, though the request had typed a hyphen.

The awkward run described the audience as “OP-1 field owners thinking about upgrading”. The Brief didn’t forbid the word. It said not to claim that OP–XY is “a superior/replacement upgrade for OP–1 field without checking an authoritative product source”, and nothing in the run’s sources described it that way. The email still opened with “OP–1 field owner, thinking about an upgrade? meet OP–XY.” The agent flagged that the framing came from the request and kept it. That’s unsubstantiated upgrade framing: the request described what the audience was thinking about, not a product relationship any source supports, and the prompt let it into the copy. The change we’d test next is one more line: “If my request implies a claim the Brief or the sources don’t support, leave it out and tell me.”

### LinkedIn feature post

LinkedIn feature post

```
Request a Rainbrand Brief for [brand] for a LinkedIn post of up to [number] words announcing [feature] to [audience], linking to [URL]. Write three versions with different opening lines, taking every fact from [source]. Before showing them to me, check each version: its word count is within the limit, product and feature names are written exactly as the brand writes them, it has no fact or number that isn’t in the source, and tell me in one sentence which version best fits the Brief’s voice guidance and why.
```

For Attio, we asked for an 80-word post announcing [workflow history and the agent log](https://attio.com/changelog/2026/workflow-history-and-agent-log) to RevOps leads. All three versions passed their checks, at 66 to 69 words. Where the Brief said Attio’s guidance doesn’t settle how to capitalise the feature names, the agent followed the changelog. The awkward run described the feature as one “which saves RevOps teams hours every week”; the agent left that out and said the changelog doesn’t support it.

The failure was quieter. Told to take every fact from the changelog, the agent reused the changelog’s sentences almost word for word, so the three versions differed mainly in their first line. The line to test next: “Use your own words; don’t copy sentences from the source.”

### Alt text

Alt text

```
Request a Rainbrand Brief for [brand] for writing alt text for the images in [folder or page]. For each image, write alt text that says what it shows and why it’s there, plainly, under [number] characters, using product names exactly as the brand writes them, and include any words that appear in the image. Give purely decorative images an empty alt. Before showing me the list, check each one against those rules, give its character count, and list any image you couldn’t describe confidently instead of guessing.
```

The test set was four images for an English article: a Blue Bottle Coffee concept ad for the US, the same ad in Japanese (both from [keeping AI ads on-brand across markets](https://www.rainbrand.com/blog/keep-ai-ads-on-brand-across-markets/)), the text-free photo behind both, and a decorative texture, with a 150-character limit. It returned an empty alt for the texture and three alt texts of 143 to 147 characters. These are the two for the ads:

| Image | Alt text it wrote |
| --- | --- |
| US ad | US version, coffee in glassware: “Bella Donovan. Raspberry, Chocolate, Molasses. Bella Donovan is our take on the classic Moka-Java profile. $23” |
| Japanese ad | Japan version, same photo: “ベラ・ドノヴァン. ラズベリー／チョコレート／モラセス. ブルーボトルコーヒーで最も人気のドリップコーヒー (Blue Bottle Coffee’s most popular drip coffee). ¥2,150（税込み）” |

The US alt text never says whose ad it is. The logo is a bottle with no words, so the agent decided there was “nothing to transcribe” and left the brand out. The Japanese one names Blue Bottle only because the name is in the ad’s own copy, which the agent copied and translated. To fit the limit, it translated only that line, and it warned that English screen readers may misread the rest and that the Japanese alt is 255 bytes if a system counts bytes. The fix to test: “Name the brand when its logo appears.”

### Check before publishing

Check before publishing

```
Request a Rainbrand Review of [asset] against [brand] for colour, typography, logo and copy. The copy should read exactly: “[copy]”. Show me each verdict with its reason. Don’t change the asset. Before you finish, tell me which verdicts need a person: every fail, with what to fix, and every uncertain, with what a person should look at.
```

A [Review](https://docs.rainbrand.com/guides/request-review) checks a captured image against the brand and reports; it doesn’t change the asset. We ran this prompt on a Graza Instagram post exported from a slide template, with its headline set in a stand-in font, EB Garamond, instead of the brand’s ITC Garamond Condensed. Colour and copy passed, and typography and logo came back uncertain. The file was unchanged afterwards. In the awkward run, the expected copy had a comma and a lowercase “s” the post doesn’t have, and the copy check failed and named both differences.

The prompt didn’t say which font the post used, so the Review couldn’t tell the stand-in from the brand font and returned uncertain rather than fail. An earlier Review of the same file, told the font, failed typography. The agent did tell the person to check the font in the design file, which would have caught it, but the fix belongs in the prompt: “Tell the Review which fonts and files you used.”

## Test each prompt before you share it

A prompt that worked once for its author isn’t ready for the team. The tests above followed the same steps:

1. Pick a real, recent task the prompt should handle, and one awkward variation: a misspelled product name, a claim the source doesn’t make, a second script, or a request that contradicts the brand.
2. Run the prompt in a fresh session with the model your team actually uses, so nothing from an earlier chat helps it.
3. Read the output against the prompt’s own check, and check the check: confirm that a removed claim really was unsupported, and a passed name really is how the brand writes it.
4. Write down what failed in the entry’s known failures, and fix the prompt or the brand, whichever was wrong. A change to the prompt is untested until you rerun both tasks.

Rerun the same tasks when the model changes, because the same prompt can behave differently on a new model, and after a significant brand change, because the guidance the prompt fetches changes what it makes. What a brand change shouldn’t need is an edit to the prompt itself, if the prompt fetches the brand rather than containing it. [How to test whether an AI agent follows your brand guidelines](https://www.rainbrand.com/blog/test-whether-ai-agents-follow-brand-guidelines/) covers building a fuller test set.

## Owners, versions and retiring prompts

Give each entry one owner, usually whoever runs that kind of asset, such as the lifecycle marketer for emails. The owner approves changes, reruns the tests and retires the prompt when the job changes.

Number the versions and keep a line on what changed and why. When someone improves a prompt in their own chat, the change should come back to the library through the owner, not live in that chat. Retire prompts nobody has used in a quarter; a short library people trust is worth more than a long one they don’t.

## FAQ

### Where should we keep a prompt library?

Anywhere with an edit history and an owner: a shared doc, a Notion database, a spreadsheet or a Git repository. Start where your team already works, and move it only when you need reviews or versioning the tool can’t give you.

### How many prompts should we start with?

Four or five, one for each asset your team makes every week, each tested before it’s shared. Demand Spring suggests 10 to 20 frequent, consequential prompts; growing to that is easier once the first few are tested and trusted.

### Can one prompt library serve several brands or clients?

Yes, if the brand is a variable. A prompt that says “Request a Brief for \[brand\]” works for every brand your agents can reach, which is how agencies keep one library across clients. Test each prompt on more than one brand before relying on that.

### Do prompt libraries work for images?

Yes, with one rule: generate the picture without text or logos, then place the real logo and set the words separately. Image models redraw logos and lettering rather than placing them. [How non-designers make on-brand assets](https://www.rainbrand.com/blog/on-brand-assets-without-a-designer/) shows a post that failed for that reason and the template that fixed it.

## More in AI tools

[All 10 AI tools articles](https://www.rainbrand.com/blog/topics/ai-tools/)

- 17 Nov 2026 17 min read

    ### [How to run creative and content operations with AI](https://www.rainbrand.com/blog/creative-and-content-operations-with-ai/)

    What creative operations and content operations cover, a content production workflow stage by stage with where AI agents fit, a stage ledger from one real request with its capacity and scope decisions, what creative automation tools do, and how to keep assets on-brand as volume grows.

- 6 Nov 2026 4 min read

    ### [How to give Claude Code and Cursor your brand, not just your design system](https://www.rainbrand.com/blog/brand-context-for-coding-agents/)

    Coding agents read your design system and instruction files, but a landing page also needs your voice, product claims, imagery and logo rules. Keep code conventions in the repo, fetch the current brand for each task, and check the page before it merges.

- 6 Nov 2026 3 min read

    ### [How to use ChatGPT for marketing and stay on brand](https://www.rainbrand.com/blog/chatgpt-for-marketing/)

    Use ChatGPT to research, draft, analyse and make campaign assets from real sources. Keep projects for campaign work, fetch current brand context through a connection, and check the deliverables before publication.

## Connect Rainbrand.

Use Rainbrand with your agents, or embed it into your product.

[Use it with your agents](https://docs.rainbrand.com/mcp/getting-started) [Integrate it with your product](https://docs.rainbrand.com/getting-started/embed-into-your-product)

### ChatGPT + Codex

1. Open the Rainbrand plugin and add it.

    [Open Rainbrand plugin](https://chatgpt.com/plugins/plugin_asdk_app_6ac6e5030ec88191a88001f6dc0628ca)

2. Connect your Rainbrand account when asked. It works in ChatGPT and in Codex.

### Claude

1. Open the Rainbrand connector in Claude and add it.

    [Open Rainbrand connector](https://claude.ai/customize/connectors?modal=add-custom-connector&connectorName=Rainbrand&connectorUrl=https%3A%2F%2Fmcp.rainbrand.com%2Fmcp)

2. Select Connect, then sign in to Rainbrand.

You can also install the [Rainbrand Skill](https://docs.rainbrand.com/mcp/getting-started#standalone-rainbrand-skill) or the [plugin](https://github.com/rainbrandai/rainbrand-skills/releases/latest/download/rainbrand.claude-plugin.zip).

### Claude Code

1. Add Rainbrand's plugin marketplace in your terminal.

    ```
    claude plugin marketplace add rainbrandai/rainbrand-skills
    ```

2. Install the Rainbrand plugin, including its Skill and MCP connection.

    ```
    claude plugin install rainbrand@rainbrand
    ```

3. Start a new Claude Code session, then use /mcp to sign in to Rainbrand.

You can also [connect with MCP](https://docs.rainbrand.com/mcp/getting-started#claude-code-installation-options) or install the [Rainbrand Skill](https://docs.rainbrand.com/mcp/getting-started#standalone-rainbrand-skill).

### Cursor

1. Add Rainbrand to ~/.cursor/mcp.json.

    ```
    { "mcpServers": { "rainbrand": { "url": "https://mcp.rainbrand.com/mcp" } } }
    ```

2. Sign in from the Cursor CLI.

    ```
    agent mcp login rainbrand
    ```

You can also install the [Rainbrand Skill](https://docs.rainbrand.com/mcp/getting-started#standalone-rainbrand-skill) or the [plugin](https://github.com/rainbrandai/rainbrand-skills/releases/latest/download/rainbrand.cursor-plugin.zip).

### API

1. Create an API key in your account settings, then list your brands.

    ```
    curl https://www.rainbrand.com/api/v1/brands -H "x-api-key: $RAINBRAND_API_KEY"
    ```

Integrating Rainbrand into your app? [Read the API docs](https://docs.rainbrand.com/api).

The brand layer for agents

Rainbrand builds and maintains shared brand context for your agents and products

[Get started](https://www.rainbrand.com/auth/sign-in) [Read the docs](https://docs.rainbrand.com/)

[View pricing](https://www.rainbrand.com/pricing/)
