If "ChatGPT, write me a post" produces a different tone every time someone different asks, the problem isn't AI — it's that your prompts live on individual laptops. Move them into a shared library and AI output becomes reproducible. Your social media operation stops being a one-person dependency.

This guide shows SMBs and agencies how to build a working social-media AI prompt library from scratch. The structure works whether you choose Notion, Airtable, or a Git repo.

TL;DR

• Treat prompts as shared assets, not personal notes.
• Standardize five fields per prompt: name, variables, model, examples, evaluation.
• Version-control everything (v1, v2, v3) and log usage for A/B improvement.
• Keep the active library to 40–80 prompts. Beyond that, split by category.
• Review monthly. Archive prompts no one uses.

Why a Library, Not a Folder of Notes

Three failure modes appear when prompts stay personal.

• Quality crashes when staff turns over. Winning prompts walk out in someone's ChatGPT history.
• Brand voice drifts. Different writers, different tones, an incoherent feed.
• Nothing compounds. No record of what was changed, by whom, when, or why.

A library structurally prevents all three. For brand voice fundamentals, see the AI brand voice guide.

The Five-Field Template

Every prompt entry includes:

Naming convention

Lock the format: Category-Purpose-Channel-Number.

• Category: `Caption`, `Hook`, `CTA`, `Carousel`, `Hashtag`, `Reply`, `Story`
• Purpose: `Promo`, `Edu`, `Story`, `UGC`, `Seasonal`, `LaunchTease`
• Channel: `IG`, `X`, `TT`, `LI`, `FB`, `LINE`
• Number: zero-padded 001–999

Example: `Hook-Edu-TT-003` — TikTok educational hook, version 3.

Input variables

Mark every changing piece as `{variable}` so non-writers can plug in values.

• `{industry}` `{product}` `{audience}` `{angle}` `{char_limit}` `{emoji_count}` `{language}` `{cta_text}` `{season}`

Standardized variables let you batch-fill from a spreadsheet later.

Recommended model

Same prompt, different model, different output. Treat one prompt = one model. When you port to another, branch a new version.

The ChatGPT vs Claude vs Gemini comparison catalogs how the three behave differently in social work.

Few-shot examples

Include 2–3 ideal outputs. Real posts that performed well beat fictional examples — your tone is in your own work.

Evaluation criteria

Define numeric pass thresholds upfront:

• Character count (e.g., 120–180)
• Hashtag count (e.g., 5–8)
• Emoji range (e.g., 0–2)
• Required CTA
• Banned phrases (no "absolutely", "guaranteed", "100%")

Library Architecture

For 40–80 prompts, this directory layout works well:

``` prompt-library/ ├── 00_README.md ├── 01_brand_voice.md # Tone definition, referenced by all prompts ├── 10_caption/ │ ├── promo/ │ ├── edu/ │ ├── story/ │ └── seasonal/ ├── 20_hook/ ├── 30_cta/ ├── 40_carousel/ ├── 50_hashtag/ ├── 60_reply/ └── 99_archive/ ```

`01_brand_voice.md` content

Maintain one brand-voice file referenced at the top of every prompt:

```

Brand Voice — [Your brand]

• Tone: friendly, never preachy
• First person: "we" (skip when too formal)
• Forbidden words: "revolutionary", "ultimate", "guaranteed"
• Preferred phrasing: "you might…", "consider…"
• Emoji budget: 0–2 per post

```

Keep this brand-voice file maintained using approaches from the AI brand voice article.

Five Real Templates

These assume the brand-voice file is loaded as system context.

Template 1: Caption-Promo-IG-001

``` You are a social media copywriter. Write one Instagram post for {industry} promoting {product}, under the constraints below.

Constraints

• 140–180 characters
• 0–2 emoji
• 5–8 hashtags grouped at the end
• Angle: {angle}
• Audience: {audience}
• One CTA at the close

Examples

(2 of your highest-performing prior posts)

Output

• Caption body
• Hashtag block

```

Template 2: Hook-Edu-TT-001

``` Write 5 TikTok hook lines for {topic} that prevent first-3-second drop-off.

• ≤ 8 words each
• Each must use a number, a question, or a counterintuitive claim
• No "best ever", "ultimate", "secret"

```

Template 3: CTA-Launch-X-001

``` Write 3 X (Twitter) launch CTAs for {product}.

• ≤ 280 characters each
• Link in last line only
• Max 2 emoji
• Audience: {audience}

```

Template 4: Carousel-Edu-IG-001

``` Design a 7-slide Instagram carousel (cover + 6 slides) about {topic}.

• Slide headline: ≤ 12 words
• Slide body: ≤ 80 characters
• Final slide: CTA = {cta_text}

```

Template 5: Reply-Customer-LI-001

``` A LinkedIn customer commented "{comment}". Reply within 100 words.

• Polite, no jargon
• One question to continue the dialogue
• No exclamation points

```

Versioning

Prompts iterate like code. Track v1 → v2 → v3 with measurable changes:

The table is a worked example, not a published benchmark. Track your own deltas in your spreadsheet.

In a Git repo, this happens naturally via PRs. In Notion/Airtable, use a "version" property and move retired prompts to `99_archive/`.

Tool Comparison

Three common platforms for the library:

Pair with a content calendar to track which prompt produced which post.

Logging Usage

To run real A/B tests on prompts, capture per-post metadata:

• Date / channel / prompt name / version / author / raw AI output / final post / KPIs (reach, saves, CTR)

A monthly summary shows which prompts hit the top 30% — those stay; the rest get refactored or archived.

Industry Starter Packs

Don't build from a blank page. Start with the 20 prompts your industry actually uses.

Cafés and restaurants

• Menu intros, seasonal items, staff intros, behind-the-scenes, regular customer stories, review highlights
• See café use case and restaurants

Beauty and salons

• Permitted before/after, style suggestions, hair-care education, stylist intros
• See beauty salon use case

E-commerce and retail

• New-product teasers, in-use scenes, staff styling, limited sales
• See e-commerce use case

Fitness

• Daily stretch, transformation stories (consented), trainer intros, nutrition tips
• See fitness use case

Freelancers and solo operators

• Project highlights, real-life work commentary, inquiry funnels
• See freelancer use case

Per-Channel Differences

Same theme, different prompt. Each platform rewards a different voice.

Cross-reference with the Instagram algorithm guide and TikTok algorithm guide when calibrating per-platform thresholds.

Common Failure Modes

Failure 1: Library balloons past 100 prompts and nobody uses it

• Fix: keep ~60 active. Anything else goes to archive.

Failure 2: Brand voice file goes stale

• Fix: quarterly review minimum. Update for new product lines or audience shifts.

Failure 3: Someone keeps using personal ChatGPT

• Fix: make "no posts outside the library" an explicit team rule.

Failure 4: No A/B logs

• Fix: post metadata is a required field, reviewed monthly.

The Monthly Review Loop

Thirty minutes a month maintains the library:

• Top 20% prompts — articulate what's working
• Bottom 20% prompts — diagnose, refactor or archive
• New candidates — propose up to three for next month
• Brand voice deltas — reflect changes in `01_brand_voice.md`

Combined with the post consistently guide, this loop stabilizes content supply.

FAQ

Q1. Isn't ChatGPT's saved-prompt feature enough?

For a single user, sure. For a team that needs versioning, sharing, and A/B testing, it's not. Use a shared knowledge tool.

Q2. How many prompts is enough?

Forty to sixty for SMBs; under 80 even for agencies. Past 100 you'll accumulate dead templates nobody uses.

Q3. Can one prompt run across multiple models?

Same prompt, different model = different output. Branch versions per model.

Q4. How often should prompts be updated?

Monthly review plus a quarterly brand-voice refresh. Update sooner when a major algorithm change lands (Instagram etc.).

Q5. Can I import the library into a dedicated SNS tool?

Most AI SNS tools have their own template systems, but if you store prompts as Markdown or JSON, importing into a tool like Adpicto is straightforward.

Next Steps

Start with five templates and a brand-voice file. Run them for a month before adding more. You'll learn what your operation actually needs faster than by trying to design the perfect library upfront. Cross-check finished templates against the principles in the captions that convert guide and let only the high-performing prompts survive.