Prompt Library Starter Pack

Page 1 — Read this first

Outcome: After customizing this pack, the learner can hand any team member a tested, on-brand prompt for 30 common business tasks — instead of leaving each person to improvise.

What this is. Thirty ready-to-use prompts covering sales, HR, finance, operations, customer support, and leadership. Every prompt uses [BRACKETED PLACEHOLDERS] you replace with your own details. Each one lists its purpose, the prompt text, and what good output looks like, plus a safety note wherever the task touches people, money, or reputation.

Who it's for. Owners, office managers, and department leads at small and mid-sized businesses who want their team using AI consistently — same quality bar, same safety habits — rather than 15 people prompting 15 different ways.

When to use it. Use this pack when you are rolling AI tools out to a team, when output quality varies wildly by person, or when you catch yourself retyping the same kind of request every week. Context: a shared library is the single cheapest upgrade to team AI use — it captures your best prompts once, so nobody starts from a blank page again.

Warnings — read before first use:

1. These prompts draft; humans decide. Nothing produced by these prompts goes to a customer, an employee's file, a regulator, or the public without a person reviewing it. The prompts themselves remind the AI of this; the reminder does not replace the review.
2. Placeholder discipline. Replace every [PLACEHOLDER] before running, and never paste real customer personal data, health information, payroll details, credentials, or confidential financials into a placeholder. Summarize or sanitize instead — see the Privacy Rules for Everyday AI Use template.
3. Verify facts and figures. AI output can state wrong things confidently. Any number, date, name, or legal/tax/HR claim gets checked against your source system or a qualified professional before use. Where prompts touch regulated topics, treat output as background only (verify current requirements before relying on it).
4. Test before you distribute. Run each prompt once with fictional data before adding it to your team library. Ten minutes of testing prevents a hundred bad drafts.

Page 2 — Completed example (fictional business)

Everything below is fictional: Cedar & Vine Catering, a 9-person catering company. Sales lead Dana just finished a tasting with a prospective client, the Harborview Community Center, and wants the follow-up email out within the hour.

Dana takes prompt S2 (Post-meeting follow-up email) from the library and fills the placeholders:

You are a sales assistant for Cedar & Vine Catering, a small catering company
whose voice is warm, unfussy, and confident.

Draft a follow-up email to Marcus Lee, events coordinator at Harborview
Community Center, after today's tasting meeting.

Context from the meeting (use ONLY these facts; do not add or embellish):
- They loved the rosemary chicken and the vegetarian mezze options
- Their event: staff appreciation dinner, approx. 80 guests, evening of [DATE
  IN ~6 WEEKS]
- Their concern: staying under a per-person budget; they are comparing two
  other caterers
- We agreed I would send a revised per-person quote by Thursday

Rules: under 150 words. Reference two specific moments from the tasting.
Restate the Thursday commitment. Do not state any prices — the quote comes
separately. Do not disparage competitors. End with one low-pressure next step.
Do not invent any facts; use [BRACKETS] for anything missing.

Format: subject line, then body, ready to paste.

What came back (sample-quality output, described): A 130-word email with the subject "Great to cook for you today, Marcus." It opened by naming the rosemary chicken and the mezze — the two details Dana supplied — thanked Marcus for the candid budget conversation, confirmed "revised per-person quote in your inbox by Thursday," and closed with "If it's easier, I'm happy to hold a tentative date for the evening of [DATE] — no obligation either way." No prices appeared, no invented menu items, no competitor mentions, one placeholder correctly left visible for Dana to fill.

Why this output passed review: every fact traced back to the prompt; the one unknown stayed bracketed instead of guessed; the tone matched Cedar & Vine's voice; and the commitment (Thursday) was one Dana had actually made. She filled the date, gave it one human read, and sent it. Total time: about four minutes.

Why it would have failed review if the AI had added "our award-winning service" (unverified claim), quoted a price (explicitly excluded), or promised a discount (never authorized). That is what the review pass is for.

Page 3 — The library (blank version, ready to copy)

Copy this whole section into your team's shared workspace. Replace [PLACEHOLDERS] when you run a prompt, not in the library itself — the brackets are what make the library reusable. A general instruction worth adding to any prompt here: "Do not invent facts; use [BRACKETS] for anything you do not know."

Sales

S1. Discovery call prep brief

• Purpose: Walk into a first call knowing what to ask.
• Prompt:

You are a sales prep assistant. Build a one-page brief for a discovery call
with [PROSPECT COMPANY], a [SIZE/INDUSTRY] business. What we sell: [ONE-LINE
OFFER]. What we know so far: [2-3 SANITIZED FACTS]. Produce: 5 discovery
questions ordered from broad to specific, 3 likely objections with one-line
responses, and 2 things NOT to bring up on a first call. Do not invent facts
about the prospect; mark assumptions as (assumption).

• Good output looks like: Questions specific to the industry given, not generic "what keeps you up at night" filler; assumptions clearly labeled.
• Safety note: Research the prospect from their own public materials; verify anything the AI asserts about them before repeating it on a call.

S2. Post-meeting follow-up email

• Purpose: Send a same-day follow-up that proves you listened.
• Prompt:

You are a sales assistant for [COMPANY], whose voice is [3 ADJECTIVES].
Draft a follow-up email to [FIRST NAME], [ROLE] at [PROSPECT], after today's
[MEETING TYPE]. Context from the meeting (use ONLY these facts): [3-5 BULLET
FACTS INCLUDING ANY COMMITMENTS MADE]. Rules: under 150 words, reference two
specific moments, restate any commitment with its deadline, no prices unless
listed above, end with one low-pressure next step. Do not invent facts; use
[BRACKETS] for anything missing. Format: subject line, then body.

• Good output looks like: See the completed example on Page 2.

S3. Objection practice partner

• Purpose: Rehearse objection handling before a big call.
• Prompt:

Role-play a skeptical [ROLE, e.g., operations director] at [PROSPECT TYPE]
evaluating [YOUR OFFER]. Your top concerns: [PRICE / TIMING / SWITCHING
COSTS — PICK REAL ONES]. Raise one objection at a time, respond to my answers
realistically — push back when I'm vague, soften when I'm concrete. After 4
rounds, break character and score my answers 1-5 with one improvement each.

• Good output looks like: Objections that sting a little; a scorecard that names vague answers instead of flattering you.

S4. Proposal outline from call notes

• Purpose: Turn messy notes into a proposal skeleton you control.
• Prompt:

Below are my sanitized notes from calls with [PROSPECT]. Build a proposal
OUTLINE only (no prose): executive summary bullets, their 3 stated problems
in their own words, proposed approach, scope in/out, timeline placeholders,
and open questions I must answer before drafting. Flag any section where my
notes are too thin. Notes: [PASTE SANITIZED NOTES]

• Good output looks like: An outline that mirrors the prospect's language and honestly flags thin sections rather than padding them.
• Safety note: Pricing, legal terms, and guarantees are written by humans and reviewed by whoever owns them — never drafted into final form by AI.

S5. Lost-deal debrief

• Purpose: Extract lessons from a loss while it is fresh.
• Prompt:

Act as a neutral sales coach. Here is what happened in a deal we lost:
[SANITIZED TIMELINE + STATED REASON]. Give me: 3 alternative explanations
beyond the stated reason, 2 questions I could ask the prospect in a graceful
close-out note, 1 process change worth testing on the next 5 deals, and the
counterargument — why this loss might NOT be a pattern. Do not assign blame
to individuals.

• Good output looks like: Hypotheses framed as testable, not verdicts; the "might not be a pattern" section taken seriously.

HR

H1. Job description first draft

• Purpose: Get from "we need someone" to a reviewable draft fast.
• Prompt:

Draft a job description for a [TITLE] at [COMPANY], a [SIZE] [INDUSTRY]
business in [LOCATION/REMOTE]. Day-to-day: [4-6 REAL TASKS]. Must-haves:
[3 MAX]. Nice-to-haves: [3 MAX]. Voice: [ADJECTIVES]. Rules: plain language,
no cliches ("rockstar," "wear many hats"), separate must-haves from
nice-to-haves honestly, use gender-neutral wording, leave pay as [PAY RANGE]
for us to complete. Format: title, one-paragraph hook, responsibilities,
qualifications, how to apply.

• Good output looks like: Requirements that match the actual job rather than an inflated wish list; neutral wording throughout.
• Safety note: A human sets the pay range and confirms requirements are genuinely necessary — inflated requirements screen out good candidates unfairly. Employment law varies by location; have your HR advisor review before posting (verify current requirements before relying on this).

H2. Structured interview question set

• Purpose: Ask every candidate the same fair, job-related questions.
• Prompt:

Create a structured interview kit for a [TITLE] role whose core duties are
[3-4 DUTIES]. Produce: 6 behavioral questions tied to those duties ("tell me
about a time..."), 2 realistic work-sample scenarios, and for each question a
short note on what a strong vs. weak answer sounds like. Exclude anything
touching age, family plans, health, religion, or other protected
characteristics. Same questions for every candidate.

• Good output looks like: Questions about the work, not the person; usable strong/weak answer notes.
• Safety note: Hiring decisions are made by people. AI never scores or ranks candidates here; it only helps you build a fair, consistent process. Have HR counsel review your kit (verify current requirements before relying on this).

H3. Policy explained in plain language

• Purpose: Turn policy legalese into something employees actually read.
• Prompt:

Rewrite the following policy excerpt in plain language for employees at a
[SIZE] company. Keep every rule intact — simplify wording, not meaning.
Format: what this policy is, what it means for you day-to-day (bullets),
what to do in common situations (2-3 examples), who to ask. Flag any sentence
where simplifying might have changed the meaning, so a human can check it.
Policy text: [PASTE POLICY EXCERPT — INTERNAL POLICY ONLY, NO EMPLOYEE DATA]

• Good output looks like: Shorter sentences, same rules, and honest flags on any meaning-sensitive simplification.
• Safety note: The original policy remains the official version; the plain-language version says so. HR approves before distribution.

H4. New-hire onboarding checklist

• Purpose: Stop reinventing week one for every hire.
• Prompt:

Build a first-30-days onboarding checklist for a [TITLE] joining [COMPANY],
a [SIZE] [INDUSTRY] business. Inputs: their manager is [ROLE], tools they'll
use: [LIST], first project: [ONE LINE]. Organize by: before day 1, day 1,
week 1, weeks 2-4. Each item gets an owner ([MANAGER]/[IT]/[HR]/[BUDDY]) and
a "done when" statement. Include 2 check-in conversations with suggested
questions.

• Good output looks like: Items with owners and completion criteria, not vague "get settled in" filler.

H5. Difficult-conversation preparation

• Purpose: Prepare a fair, calm structure for a hard talk.
• Prompt:

Help me prepare (NOT script) a difficult conversation with an employee about
[TOPIC, e.g., repeated missed deadlines — NO NAMES]. Facts I can document:
[2-4 SPECIFIC, SANITIZED FACTS]. Produce: a 4-part conversation structure
(open, facts, listen, agree next steps), 3 questions that invite their side,
2 phrases to avoid and why, and a reminder list of what I should NOT commit
to in the moment. Tone: direct and respectful.

• Good output looks like: A structure that spends as much time listening as talking; no diagnosis of the employee's motives.
• Safety note: No names or identifying details in the prompt. Anything that could lead to discipline or termination involves HR guidance from a person, and is documented by a person. The AI prepares you; it does not judge the employee.

Finance

F1. Variance explanation narrative

• Purpose: Turn a numbers gap into a readable explanation for owners.
• Prompt:

Write a variance summary for [MONTH]. Data (verified, sanitized): budgeted
[CATEGORY] = [AMOUNT], actual = [AMOUNT], drivers we identified: [2-3 KNOWN
REASONS]. Audience: owners who read fast. Rules: 150 words max, lead with the
one number that matters, attribute variance ONLY to the drivers listed —
if drivers don't fully explain the gap, say "unexplained: [AMOUNT]" rather
than guessing. Format: headline, 3 bullets, one recommended question for
next month.

• Good output looks like: The unexplained remainder stated honestly instead of papered over.
• Safety note: The AI narrates numbers you verified; it never calculates or reconciles them. Pull figures from your accounting system, not memory.

F2. Payment reminder sequence

• Purpose: Chase receivables consistently without souring relationships.
• Prompt:

Draft a 3-email payment reminder sequence for [COMPANY], voice: [ADJECTIVES].
Invoice context: [CLIENT TYPE — NO REAL NAMES], invoice [NUMBER PLACEHOLDER],
amount [AMOUNT PLACEHOLDER], terms [NET-X]. Email 1 (due date +5 days):
friendly nudge. Email 2 (+20): firm, offers a payment conversation. Email 3
(+40): final notice before we pause work — professional, no legal threats.
Each under 120 words with subject line. Do not invent late-fee terms; use
[LATE FEE TERMS] placeholder.

• Good output looks like: Three genuinely different temperatures; escalation in clarity, not hostility.
• Safety note: A human confirms amounts against the ledger before each send — chasing an already-paid invoice costs more goodwill than it recovers. Late-fee and collections language must match your actual contract terms.

F3. Budget assumptions checklist

• Purpose: Stress-test a budget before you commit to it.
• Prompt:

Act as a skeptical CFO reviewing a [YEAR] budget for a [SIZE] [INDUSTRY]
business. Our headline assumptions: [3-5 ASSUMPTIONS, e.g., "revenue +10%,"
"hire 2 techs in Q2"]. For each: the question a lender would ask, what
evidence would support it, and the early warning sign it's failing. Then
list 3 assumptions we are probably making WITHOUT realizing it. Do not
predict the economy; interrogate our logic.

• Good output looks like: Uncomfortable questions and observable early-warning signs, not forecasts.
• Safety note: Output is a thinking aid. Budget decisions rest with owners and your accountant (verify current figures and tax treatment before relying on this).

F4. Vendor quote comparison

• Purpose: Compare quotes on more than the bottom-line price.
• Prompt:

Compare these [N] vendor quotes for [PURCHASE]. Sanitized details per vendor:
[VENDOR A: price, terms, timeline, what's included/excluded — repeat per
vendor, letters not names]. Build a comparison table with rows: upfront cost,
ongoing cost, what's NOT included, timeline, switching/exit terms, open
questions to ask. Then list which differences actually matter for a company
our size ([SIZE]) and which are noise. Recommend the 3 clarifying questions
to ask before choosing — do NOT recommend a vendor.

• Good output looks like: An exclusions row that surfaces the hidden costs; no premature winner declared.
• Safety note: The human makes the pick. Verify every quoted figure against the actual quote documents — transcription errors in, wrong decision out.

F5. Month-end close checklist

• Purpose: Document your close process so it survives vacations.
• Prompt:

Draft a month-end close checklist for a [SIZE] [INDUSTRY] business using
[ACCOUNTING TOOL]. Our close today involves roughly: [LIST WHAT YOU DO,
ROUGH ORDER]. Organize into: daily-if-possible, first 3 business days,
by day 10. Each item: action verb, owner placeholder [ROLE], and "done when"
check. Add a "common errors to check" section based on the steps I listed.
Mark any step you added that I didn't mention with (suggested) so I can
verify it applies to us.

• Good output looks like: Your real process, tightened — with suggested additions clearly labeled, not smuggled in.

Operations

O1. SOP from rough steps

• Purpose: Turn "how Maria does it" into a document anyone can follow.
• Prompt:

Turn these rough steps into a standard operating procedure. Task: [TASK].
Who performs it: [ROLE]. Rough steps as told to me: [PASTE MESSY STEPS].
Format: purpose (1 line), when triggered, tools needed, numbered steps
(one action each, no step over 2 sentences), quality check at the end,
what to do when it goes wrong ([2-3 KNOWN FAILURE POINTS]), owner and
review date placeholders. Flag any gap where the steps skip something a
first-timer would need.

• Good output looks like: Steps a brand-new hire could follow; honest flags where your rough notes had gaps.
• Safety note: The person who actually does the task reviews the SOP before it becomes official — AI cannot know your shop floor.

O2. Meeting notes to action items

• Purpose: End the "wait, who was doing that?" cycle.
• Prompt:

Extract action items from these meeting notes (names replaced with
initials). Rules: only items with a clear owner and action — list vague
intentions separately under "Discussed, no owner." Format: table with
columns Action / Owner / Due / Blocked by. Then one line: the single
decision from this meeting that affects the most people. Do not invent
owners or dates that aren't in the notes. Notes: [PASTE SANITIZED NOTES]

• Good output looks like: A short honest table plus a "no owner" list that shames the right things.

O3. Five-whys process breakdown

• Purpose: Find the process cause behind a recurring mistake.
• Prompt:

Facilitate a five-whys analysis. The problem: [ONE SENTENCE, e.g., "we
missed two scheduled service appointments last week" — NO EMPLOYEE NAMES].
Known facts: [3-5 FACTS]. Ask me "why" one round at a time, wait for my
answer, and push back if my answer blames a person rather than a process.
After 5 rounds, summarize: root cause hypothesis, cheapest test of that
hypothesis, and the process change IF the test confirms it.

• Good output looks like: The AI actually pauses each round and redirects blame-answers toward process gaps.
• Safety note: Root-cause work is about systems, not fault. Keep names out of prompts; keep discipline decisions with humans entirely.

O4. Vendor escalation email

• Purpose: Escalate a vendor problem firmly without burning the bridge.
• Prompt:

Draft a vendor escalation email. Context: [VENDOR TYPE] has [ISSUE, e.g.,
"delivered late 3 of the last 4 orders"]. Verified facts: [DATES/ORDER
NUMBERS AS PLACEHOLDERS]. Relationship: [LENGTH, VALUE TO US]. Rules: under
180 words, facts before feelings, one specific ask with a date, one
consequence stated plainly (e.g., "we'll need to qualify a second supplier"),
no threats we won't keep, professional throughout. Format: subject, body,
and a one-line internal note on what to do if there's no reply by [DATE].

• Good output looks like: Specific, dated, calm; a consequence you would actually follow through on.
• Safety note: Verify every date and order number first. If contracts or legal remedies come into play, that email is written with counsel, not a chatbot.

O5. Training one-pager from an SOP

• Purpose: Compress an SOP into the sheet that gets taped to the wall.
• Prompt:

Compress this SOP into a one-page training aid for [ROLE]. Keep: trigger,
the numbered steps as short imperatives, the quality check, and the top 2
mistakes with their fixes. Cut: background, history, edge cases (link them
as "see full SOP"). Reading level: new hire on day 2. Format: fits one page,
scannable in 60 seconds. SOP text: [PASTE SOP]

• Good output looks like: Imperatives ("Check valve. Log reading.") not paragraphs; nothing on the page a day-2 hire doesn't need.

Customer support

C1. Frustrated-customer reply draft

• Purpose: Answer heat with competence, fast.
• Prompt:

Draft a reply to a frustrated customer. Their message (sanitized, no
personal details): [PASTE MESSAGE WITH NAMES/ACCOUNT INFO REMOVED]. What
actually happened, verified: [FACTS]. What we can offer: [REMEDY WE'VE
DECIDED — the human decides this, not you]. Rules: acknowledge specifics
of their complaint in the first two sentences, no defensiveness, no
company-policy recitals, state the remedy and the exact next step with a
time, under 150 words. Do not promise anything beyond the remedy listed.

• Good output looks like: The customer's actual issue mirrored back accurately; one remedy, one next step, one timeframe.
• Safety note: The remedy decision is made by a human before prompting — the AI words the offer; it never chooses it. Strip names, order numbers, and contact details before pasting; re-add them after review.

C2. Ticket triage rules draft

• Purpose: Sort incoming requests consistently, even on busy days.
• Prompt:

Help me build triage rules for incoming [CHANNEL: email/phone/form]
requests at a [INDUSTRY] business. Our categories: [e.g., URGENT-SAFETY /
SERVICE REQUEST / BILLING / GENERAL]. For each: 5 example phrases customers
actually use, the response-time target [OUR TARGETS], and who owns it
[ROLE]. Then give me the 3 hardest judgment calls between categories and a
tiebreaker rule for each. Format: one table plus the judgment-call list.

• Good output looks like: Realistic customer phrasing (not textbook language) and tiebreakers a stressed human can apply in 5 seconds.
• Safety note: Anything involving safety, legal threats, or medical issues routes to a human immediately — put that rule above the table, in bold, in your final version.

C3. Knowledge-base article from a resolved ticket

• Purpose: Solve it once, publish it forever.
• Prompt:

Turn this resolved support case into a customer-facing help article.
Case summary (sanitized — no customer details): problem [PROBLEM],
cause [CAUSE], fix [STEPS THAT WORKED]. Audience: customers who are
[NON-TECHNICAL/TECHNICAL]. Format: title phrased as the customer's
question, 2-line answer up top, numbered steps with what the customer
should see after each, "if this didn't work" section pointing to [CONTACT
CHANNEL]. Rules: no internal jargon, no blame, under 300 words.

• Good output looks like: A title someone would actually search; steps with visible checkpoints ("you should now see...").

C4. Service-failure apology with make-good

• Purpose: Own a real mistake in writing, correctly.
• Prompt:

Draft an apology email for a service failure. What happened, verified:
[FACTS — WHAT WE GOT WRONG, SANITIZED]. Impact on customer: [IMPACT].
Make-good we have approved: [SPECIFIC REMEDY + WHO APPROVED IT]. What we
changed so it doesn't recur: [REAL CHANGE, OR OMIT THIS LINE]. Rules:
apologize once, specifically, in the first sentence; no "we apologize for
any inconvenience"; no excuses; no admissions beyond the listed facts;
under 140 words. Format: subject line + body.

• Good output looks like: A specific first-sentence apology, the approved remedy, and — only if true — the fix. Nothing invented.
• Safety note: Apologies can carry legal weight. A manager approves every service-failure email before send; if injury, significant loss, or liability is involved, it goes through counsel instead of this prompt.

C5. Feedback theme summary

• Purpose: Turn a month of comments into three findings leadership will read.
• Prompt:

Analyze this batch of customer feedback (all identifying details removed).
Produce: top 3 themes with a representative quote each and rough frequency
(count per theme, note "small sample" if under 30 items), 1 emerging theme
too small to rank, 1 thing customers praise that we should protect, and 2
questions the data raises but cannot answer. Do not propose solutions —
findings only. Feedback: [PASTE ANONYMIZED FEEDBACK]

• Good output looks like: Honest counts with sample-size caveats; quotes that sound like real customers; questions listed instead of overreach.
• Safety note: Anonymize before pasting. Counts from AI are approximate — spot-check a theme by hand before repeating its frequency to leadership.

Leadership

L1. Weekly team update from bullets

• Purpose: Ship a consistent Friday update in 10 minutes.
• Prompt:

Turn my bullets into our weekly team update. Voice: [ADJECTIVES — e.g.,
direct, warm, no corporate-speak]. Structure every week: Wins / In motion /
Stuck (and what would unstick it) / Next week's one priority / Shout-out.
Rules: under 250 words, keep my facts exactly as given, no motivational
padding, if a section has no content write "none this week" rather than
inventing. My bullets: [PASTE BULLETS]

• Good output looks like: Your facts, your voice, same skeleton every week — the consistency is the feature.

L2. Decision memo (options and tradeoffs)

• Purpose: Force a fuzzy decision into a one-page structure.
• Prompt:

Structure a decision memo. Decision to make: [ONE SENTENCE]. Deadline:
[DATE]. Options we see: [OPTION A / B / C — ONE LINE EACH]. What we know
(verified): [FACTS]. What we don't know: [UNKNOWNS]. For each option:
strongest case for, strongest case against, what would have to be true for
it to win, reversibility (easy/hard to undo). End with: the information
most worth getting before deciding, and the cost of deciding late. Do NOT
recommend an option.

• Good output looks like: Symmetrical steelmanning of every option; the "what would have to be true" lines doing real work.
• Safety note: The memo structures thinking; the decision belongs to the accountable human. For decisions affecting jobs, money movement, or legal exposure, add professional advice to the inputs.

L3. Sensitive-change announcement draft

• Purpose: Announce hard news without spin or corporate fog.
• Prompt:

Draft an internal announcement about [CHANGE, e.g., "reorganizing the
service team" — NO NAMES, NO UNANNOUNCED PERSONNEL DETAILS]. Facts I can
share now: [APPROVED FACTS ONLY]. What I cannot address yet: [TOPICS].
Rules: lead with the change and the reason in the first 3 sentences, no
euphemisms ("rightsizing"), acknowledge what's hard about it in one honest
sentence, state exactly where and when people can ask questions, under 200
words. Flag any sentence that might overpromise so I can review it.

• Good output looks like: News in sentence one, a reason people can repeat accurately, and a real Q&A channel.
• Safety note: Anything touching roles, pay, or employment status gets HR and (where relevant) legal review before sending — no exceptions, no matter how good the draft reads. Never put unannounced personnel decisions into an AI tool.

L4. Red-team my plan

• Purpose: Pay an AI to argue with you before reality does.
• Prompt:

Here is a plan I'm confident about: [PASTE PLAN SUMMARY, SANITIZED].
Red-team it. Produce: the 3 most likely failure points and the early
signal of each, the weakest assumption stated as a question, who bears
the cost if this fails (customers/staff/cash), the strongest version of
"do nothing instead," and one cheap test to run within 2 weeks before
committing fully. Be blunt; do not balance criticism with praise.

• Good output looks like: Criticism that stings and early-warning signals you can actually watch.

L5. Meeting agenda and facilitation plan

• Purpose: Make the meeting earn its calendar slot.
• Prompt:

Design a [LENGTH]-minute meeting. Purpose (decision or update?):
[ONE SENTENCE]. Attendees by role: [ROLES]. Decisions needed: [LIST].
Build: timed agenda with a purpose line per block, the one pre-read to
send (and what it must contain), 2 questions to draw out quieter
attendees, and a closing checklist (decisions restated, owners, dates).
If this could be an email instead of a meeting, say so and draft that
email instead.

• Good output looks like: Timeboxes that add up, and the honesty to tell you it should have been an email.

Customization worksheet (blank)

Complete this once, as a team, before distributing the library. Paste the answers into a "house rules" note at the top of your library.

Practice (do this before rollout): Have each department pick its single highest-volume task, run the matching prompt with fictional data, and bring the output to a 20-minute review meeting. Keep what passed, fix what didn't, delete what nobody will use. A 12-prompt library people trust beats a 30-prompt library people ignore.

Final page — Review checklist and sign-off

Review checklist — before the library goes live:

• [ ] Every prompt tested at least once with fictional data
• [ ] Voice samples collected and sanitized (no customer or employee data)
• [ ] House rules note completed from the customization worksheet
• [ ] Human-approval list agreed and written at the top of the library
• [ ] Never-paste list adopted from the Privacy Rules template and linked
• [ ] Safety notes read by every department lead (HR and Finance leads read theirs twice)
• [ ] Pilot run with 3 users for one week; feedback collected
• [ ] Baseline measured so value is provable: minutes per draft on 2-3 common tasks, before vs. after (use your numbers, not vendors' claims)
• [ ] Owner and review date assigned below

Next step: Once your team runs on a shared library, the pattern becomes obvious: the same five prompts, dozens of times a week. That is the moment to learn what comes after prompting — Understanding AI Agents explains how repetitive prompt-driven work becomes governed automation, and what boundaries keep it safe. When you want those high-volume workflows built with audit trails and human approval gates, Epic Dreams designs them on Ed OI — start at /services.