Shared Prompt Library: Structure, Governance, 80/20 Starter Set

After watching 30+ SMBs try to scale AI use beyond a handful of power users, my conclusion is simple: the teams that stick are the ones with a shared prompt library treated like code, not like a Slack thread. The ones that don't stick keep reinventing the same prompt every Tuesday.

Why does a prompt library matter?

Because without one, every employee is solving the same five problems privately, badly, at 4 PM on Friday.

A solo prompt is an idea. A shared library is institutional memory. When a new hire joins on Monday and finds a tested "follow-up after a discovery call" prompt with a one-line note explaining when to use it — they're productive on day one instead of day fifteen.

Definition: Prompt library — a versioned, governed collection of reusable prompts, organized by role and task, with metadata about owner, last-tested date, and known failure modes. Not a Notion page with copy-pasted screenshots.

The library is also the only durable artifact that survives tool changes. Switch from ChatGPT to Claude next year? The prompts mostly transfer. Lose a key employee? Their best workflows stay.

What folder shape actually works?

Two axes: role family on top, task type underneath. Resist the urge to add a third axis ("by department"). Resist twice.

/library
  /sales
    /outbound-email
    /discovery-prep
    /follow-up
    /proposal-review
  /customer-success
    /onboarding-message
    /renewal-talk-track
    /escalation-summary
  /finance
    /ap-coding
    /variance-explanation
    /forecast-narrative
  /marketing
    /blog-outline
    /case-study-draft
    /campaign-brief
  /people
    /jd-rewrite
    /screening-rubric
    /onboarding-week-one
  /ops
    /meeting-notes
    /weekly-update
    /sop-draft
  /_shared
    /redaction-checklist
    /tone-and-voice
    /quality-self-check

Six role families, three to four task types each, a shared folder for cross-cutting helpers. About 20-30 prompts at launch is the right scope — enough to be useful, not so much that nobody can find anything.

Definition: Cross-cutting prompt — a prompt that applies regardless of role (redact PII before paste, tone-shift to formal, check this output for three common failure modes). Lives in _shared.

What does each prompt entry need?

Five fields, every entry, no exceptions.

1. Title — verb + noun, plain English. "Draft discovery follow-up email."
2. When to use it — one sentence. "After a 30-min discovery call where the prospect named 2+ pain points."
3. The prompt itself — copy-pasteable, with [BRACKETED PLACEHOLDERS] for variable inputs.
4. Owner — a person, not a team. Owners are responsible for testing twice a quarter.
5. Last tested — date plus model used. If older than 90 days, the prompt shows a "needs re-test" flag.

Optional but high-value: known failure modes ("doesn't handle multi-stakeholder calls well"), and one example input/output pair.

Copy/paste prompt entry template

This is the entry format we hand teams as their template. Adapt the fields, keep the structure.

# [VERB + NOUN TITLE]

**When to use:** [one sentence]
**Owner:** [name]
**Last tested:** [YYYY-MM-DD on MODEL]
**Known failure modes:** [list, one line each]

---

## Prompt

You are a [ROLE] assistant for a [COMPANY TYPE].

CONTEXT:
- [Static context line 1]
- [Static context line 2]

INPUTS (paste below):
- [INPUT 1 — what the user pastes]
- [INPUT 2 — optional]

INSTRUCTIONS:
- [Step 1]
- [Step 2]
- [Constraint: tone, format, length]
- [Anti-instruction: what NOT to do]

OUTPUT FORMAT:
- [Exact shape — JSON, markdown, plain]
- [Length limit if any]

---

## Example

INPUT:
[realistic example input]

OUTPUT:
[the output we want, lightly edited if needed]

The Example block is the part most libraries skip. It's also the part that turns "I tried it and got weird output" into "I see what shape it expects."

Tool tip (Course for Business): In our 6-week program, building the library is week two, and it's done with the AI Champions (1:15-20) ratio sitting beside each role family. The Augment, don't replace rule applies here: the champion doesn't write the prompts in a vault — they write them with the salesperson, controller, or recruiter who'll use them. That's the difference between a library that has 200 prompts and 4% reuse, and one with 30 prompts and 70% reuse. Walk through the program at https://course.aiadvisoryboard.me/business.

How do you keep the library alive?

Three lightweight governance rules.

1. Twice-a-quarter test, by the owner. If a prompt isn't tested in 90 days, it's flagged. If it's not tested in 180 days, it's archived. Archive, not delete — sometimes the old version helps debug a new failure.
2. Add-only-after-second-use. A new prompt enters the shared library only after a second person has used it successfully. Solves the "I made this once and added it" graveyard problem.
3. One owner per folder. A named human, not a Slack channel. Owner reviews additions, runs the test cadence, deprecates dead prompts.

That's it. Anything more bureaucratic kills contribution.

Definition: Add-only-after-second-use rule — a prompt is admitted to the library only after a second team member has used it on a real task and confirmed it produced acceptable output. Filters out enthusiasm-only contributions.

What's the 30-prompt starter set?

Six prompts per role family across five families, plus five cross-cutting helpers. Targets the top-volume task in each family.

• Sales: outbound cold email, discovery prep brief, follow-up after call, proposal QA pass, account research summary, objection-handling draft.
• CS: onboarding welcome, check-in template, escalation summary, renewal talk track, churn-risk note, knowledge-base article draft.
• Marketing: blog outline, case study draft, campaign brief, social post variants, ad copy A/B pairs, newsletter summary.
• Finance/Ops: invoice coding, variance explanation, forecast narrative, vendor email, meeting notes, weekly status update.
• People/HR: JD rewrite, screening rubric, interview question set, onboarding plan, performance feedback draft, exit interview summary.
• Cross-cutting: PII redaction checklist, formal tone shift, output quality self-check, summarize-to-three-bullets, "explain this like I'm a non-expert."

Thirty prompts. Two weeks of focused work for an AI champion plus a representative from each role family. That's it.

Team scan (what AI champions report after week 1)

• Most-used prompt in week one: meeting notes / action items extraction (every department uses it).
• Highest-leverage role family for library investment: sales (used most, scales fastest, ROI most visible).
• Reuse rate of starter set after 30 days: typically 60-75% — solid evidence the taxonomy held.
• Failure pattern: prompts written without an example block; reuse drops by half.
• Failure pattern: 200-prompt libraries assembled in week one; people can't find anything; usage collapses.
• One AI champion per ~17 staff owns the library; folder owners are the day-to-day stewards.
• First friction: people want to add their pet prompt without the add-only-after-second-use rule; hold the line.
• First win: a sales rep finds "follow-up after a discovery call" on day two and ships their first AI-drafted email within an hour.
• Quality-gate signal: prompts last tested within 60 days have ~3× the reuse of those last tested 90+ days ago.
• Library survives a tool change: when one team migrated from one provider to another, 90% of prompts ported with minor edits.

Micro-case (what changes after 7-14 days)

A 90-person B2B SaaS company shipped a 28-prompt starter library in twelve working days. Week one: the AI champion sat with one rep from each role family and drafted entries; week two: each entry was tested by a second person and either admitted or rejected. Six prompts were rejected (too narrow, too generic, or relied on undocumented context). After two weeks, reuse rate across the library hit roughly 65%, and the most-used prompt (meeting notes extraction) was running 40+ times a day across the company. The CEO's signal that it had worked: a new joiner in customer success used three library prompts on day three without anyone walking her through them.

Note on this case: This example is illustrative — based on typical patterns we observe with companies of 30-500 employees, not a single named client. Specific numbers are rounded approximations of common ranges, not guarantees.

Tool tip (Course for Business): The single biggest trap with prompt libraries is treating them as a one-time documentation project. They're not — they're a living system that needs an owner, a test cadence, and a contribution rule. Our 6-week program builds the library in week two using Shoulder-to-Shoulder hot seats with each role family, then hands over the governance rules and one named owner per folder. Augment, don't replace also applies here: the library belongs to the people who use it, not to an "AI team." Book a 30-min mapping call at https://course.aiadvisoryboard.me/business.

FAQ

Where should the library actually live? Anywhere your team already reads. Notion, Confluence, a GitHub repo, a shared Drive folder. The wrong choice is "a new tool nobody opens." Pick the platform with the highest existing traffic and the simplest copy-paste UX.

Should the prompts be private to each user? No — the whole point is shared. Private prompt collections are how organizations end up with 12 versions of the same outbound email prompt and zero institutional memory.

What about prompt versioning? Lightweight is fine for SMBs — keep the previous version in a comment block or a _archive subfolder when you change a prompt materially. Heavy version control (semver, change logs per prompt) is overkill at this size.

Do we need a tool for this, or is markdown enough? For 30-300 person SMBs, structured markdown in a place your team already reads is enough. Dedicated prompt-management tools are useful at higher scale or when you're shipping prompts into production agents — different problem, different tool category.

Conclusion

The prompt library is the lowest-cost, highest-leverage piece of AI infrastructure an SMB can ship. Three things make or break it: a clean two-axis taxonomy, an entry template that includes an example, and three lightweight governance rules people will actually follow.

Block two weeks. Build the 30-prompt starter set with one champion and one user per role family. Ship the governance doc the same day you ship the library.

If you want every employee to ship their first AI automation in five days — book a 30-min call and we'll map your team's first week at https://course.aiadvisoryboard.me/business.