Templates · 14 copy-and-use · No gate · updated 1 November 2026

The prompt and template library for your first AI agent

A first AI agent needs six written things before any code: a ten-line rulebook saying what it may do alone, with approval, and never; an escalation list of the messages that go straight to a person; the WhatsApp templates it will send, written as utility messages so they are cheaper and deliverable; a one-page data-flow diagram; a baseline measured for two weeks before it goes live; and a checklist to hold any vendor, including us, to. The templates below are those six, filled in enough to copy and adapt in an afternoon, plus three prompts that make an AI assistant draft the first version for you.

Fourteen templates we use on every agent build, free and ungated: the ten-line rulebook, the escalation list, five WhatsApp templates written to pass Meta’s category rules, three scoping prompts to run with any AI assistant, the twelve-question vendor guardrails checklist, and the pilot baseline sheet. Copy, fill the brackets, keep. They are the artefacts the pilot produces anyway; having them first makes the scope call shorter.

How to use these

Each template is plain text with the parts you change in [square brackets]. Copy it, fill the brackets, and keep it in whatever your business keeps documents in. None of them needs a developer; all of them make a developer faster. The three prompts at the end are for running with any AI assistant to produce a first draft of the rulebook, the escalation list and the data-flow diagram from a description of your business.

They are published ungated because they are more useful as a shared vocabulary than as a lead magnet: when a clinic arrives at the scope call with a rulebook and an escalation list, the first hour of the pilot is already done, and the pilot is the thing we sell.

The rulebook and the escalation list

The two documents every agent is built against. Ten lines and a list; if either takes more than an hour, the workflow is not understood well enough yet.

The ten-line agent rulebook

Before any agent is scoped. One rulebook per workflow.

AGENT RULEBOOK — [workflow name], v1, [date], owner: [name]

1. PURPOSE: This agent [books / qualifies / chases / reconciles] [what] for [who], on [channel].
2. TRIGGER: It starts when [a message arrives on / a row appears in / the clock reaches] [source].
3. MAY DO ALONE (D = draft, A = act with approval, X = act alone):
   - [action, e.g. answer hours, fees, location from the approved facts list] — X
   - [action, e.g. offer available slots and book] — A
   - [action, e.g. send the reminder template] — X
4. MUST NEVER: [give medical/legal/financial advice]; [quote a price not in the system]; [promise an outcome]; [message anyone without consent on record].
5. HANDS TO A PERSON WHEN: anything on the escalation list; any question not in the approved facts; any message it is less than sure about; anyone who asks for a human.
6. SOURCES OF TRUTH: facts from [approved facts list]; availability and prices from [system]; nothing from general knowledge.
7. TONE: [short, warm, first name, the client's language]. Says it is an automated assistant in the first message.
8. LOGGING: every message, tool call and decision logged against [the lead / the booking / the matter], readable by [who] without a developer.
9. OFF SWITCH: [who] can turn it off at [where]; all traffic then goes to [person/number].
10. DONE LOOKS LIKE: [the booking is in the system and confirmed] / [the checklist is complete] / [the summary is in the CRM] and a person was involved only where lines 3 and 5 say so.

Most first agents ship with no X on an action that touches money or a customer’s record. Promote a line from A to X only after two weeks of approvals with no corrections.

The escalation list starter

Alongside the rulebook. Read your last fifty inbound messages before you call it complete.

ESCALATION LIST — [workflow], hands to [person / number] immediately, with the whole thread

ALWAYS, ANY INDUSTRY
- Asks for a human, or repeats the question a second time
- Anger, distress, a threat, a complaint about a person
- Anything legal: notice, lawyer, court, refund demand, "I will report"
- Anything about money beyond what the system shows: discount, negotiation, dispute
- Anything the agent is not sure about

HEALTHCARE (add if clinic/dental)
- Any symptom, pain, bleeding, medication, dosage, "is it serious"
- Anything about a child under [age], pregnancy, or an emergency word

PROFESSIONAL SERVICES (add if CA/law firm)
- "What should I do", "will I win", "is this legal", any request for an opinion
- A deadline, limitation date, notice date or hearing within [7] days
- The other party's name matches an existing client (conflicts)

PROPERTY / HOSPITALITY (add if brokerage/hotel)
- Questions about the "kind of people" in a locality, community, religion, safety
- A guest complaint during a stay; anything about an injury or a safety hazard

HOW TO ESCALATE
- Acknowledge in the client's language: "I'm passing this to [name], who will reply by [time]."
- Stop all automated messages to this contact until [person] closes the issue.

Ten items is normal. Two items means you have not imagined the hard messages yet.

WhatsApp Business templates that pass review

Written as utility messages where Meta allows it (cheaper, more deliverable than marketing), with {{n}} placeholders in Meta’s format. Submit through your provider; approval usually takes minutes to a day.

Appointment reminder (utility)

The evening before. Reply options let the agent update the schedule.

Hi {{1}}, a reminder of your appointment at [Business name] on {{2}} at {{3}} with {{4}}.
Reply 1 to confirm, 2 to reschedule, or call [number] if anything changes.
[Business name], [address line / map link]

Keep the business name in the body; it is what makes it a utility message about an existing appointment rather than marketing.

Recall / overdue follow-up (utility)

For patients, members or clients who are due for something they already have a relationship with you for.

Hi {{1}}, this is [Business name]. Your {{2}} was due around {{3}}.
Would you like us to book it? Reply with a day that suits you, or 1 for the next available slot.
Reply STOP to opt out of reminders.

The STOP line is what keeps a recall message on the right side of consent rules, and the agent must honour it on the spot.

Document request (utility)

CA firms, law firms, lenders, anyone with a checklist. One message per request, not a list of ten.

Hi {{1}}, for {{2}} we still need: {{3}}.
You can send it here as a photo or PDF, or upload at {{4}}.
Needed by {{5}} so we can file on time. Reply here if anything is unclear. — [Firm name]

The agent fills {{3}} with only what is still missing, which is why the request is one item rather than the whole checklist again.

First reply to a new enquiry (service window, free-form)

Within the 24-hour service window after an inbound message no template is needed; this is the opening the agent uses.

Hi, thanks for messaging [Business name]. I'm the automated assistant; I can answer about [hours / fees / availability] and book you in, and a person from the team sees everything I do.
To help quickly: are you asking about [option A], [option B], or something else?

Saying it is an automated assistant in the first line is the posture every Indian advisory since March 2024 has asked for, and it lowers complaints.

Review request (utility, post-service)

The day after a completed visit, stay or delivery, and never to someone with an open complaint.

Hi {{1}}, thank you for visiting [Business name] on {{2}}. If you have a minute, a short review helps others find us: {{3}}
If anything was not right, reply here and [name] will call you.

The second line matters: it routes unhappy guests to you instead of to the review form, which is the honest version of reputation management.

Scoping prompts to run with any AI assistant

Paste one into Claude, ChatGPT or Gemini with your details filled in. They produce first drafts of the documents above; you edit, they do not decide.

Prompt: draft my rulebook from a description

When you have a workflow in your head and nothing on paper.

You are helping a small business write a ten-line rulebook for an AI agent. Do not propose features; produce the rulebook only.

Business: [type, size, city]. Channel: [WhatsApp / web chat / email].
Workflow: [describe in 3–5 sentences what happens today, who does it, what systems are involved].
What the agent must never do: [list].
Systems it can read: [list].

Produce the rulebook with exactly these headings: PURPOSE, TRIGGER, MAY DO ALONE (mark each action D/A/X and default to D or A), MUST NEVER, HANDS TO A PERSON WHEN, SOURCES OF TRUTH, TONE, LOGGING, OFF SWITCH, DONE LOOKS LIKE. Keep each line under 25 words. Then list three questions you would need answered before this could be built.

Prompt: build my escalation list from real messages

When you can paste twenty to fifty real inbound messages (names removed).

Below are real customer messages received by a [type of business] on [channel], with names and numbers removed.

For each message, say whether an automated assistant should (a) answer from an approved facts list, (b) act with a person's approval, or (c) hand to a person immediately, and why in five words.
Then produce an ESCALATION LIST: every pattern that got (c), grouped, phrased as rules a non-technical owner can read. Flag any message that involves health, law, money disputes, children, safety or distress even if it looks routine.

Messages:
[paste]

Prompt: draft my one-page data-flow diagram

Before a vendor conversation, or when a customer asks where their data goes.

Write a one-page data-flow description for an AI agent, in plain language a customer or a regulator could read.

Agent: [what it does]. Channel: [WhatsApp etc.]. Systems it reads: [list]. Systems it writes to: [list]. Model provider: [name, and whether the business plan excludes training on our data]. Country: India.

Cover: (1) what personal data is seen, field by field; (2) where each field goes and in which system it is stored; (3) who in the business can see it; (4) how long it is kept and how it is deleted on request; (5) what is logged; (6) the one switch that turns the agent off. End with the three sentences we would put in our customer-facing privacy notice about this agent.

The vendor guardrails checklist and the pilot baseline

Twelve questions to ask any agent vendor, including us, and the sheet that turns a pilot into a measured result.

The 12-question agent guardrails checklist

In every vendor conversation, before the price. A vendor who cannot answer six of these is selling a demo.

AGENT GUARDRAILS CHECKLIST — ask every vendor

1. Show me the rulebook the agent runs on. Can I read and change it without you?
2. What happens when a customer asks something not in the approved facts? Show me the exact reply.
3. Show me the escalation list and a transcript of an escalation actually happening.
4. Where is every conversation and tool call logged, and can my manager read it without a developer?
5. Where is the off switch, who can press it, and where does traffic go when it is off?
6. Which model provider, which plan, and does that plan train on our customers' data? Show me the clause.
7. Draw the data flow on one page: what is read, where it is stored, who sees it, how it is deleted.
8. What does the agent do with a message about health / law / money / a child / a threat? (Pick the ones that apply.)
9. How do we measure it? What is the baseline, taken how, over what period, and what number decides whether we continue?
10. What do we own at the end: prompts, code, accounts, data? If you disappear, does it keep running?
11. What are the running costs per month at our volume, billed to whom, and what makes them go up?
12. What has gone wrong on a previous deployment, and what did you change afterwards?

Our own answers are on the how-we-work and pilot pages. If a vendor’s answers are better than ours, hire them.

The pilot baseline sheet

Two weeks before the agent goes live, and the same two weeks after. One row per lead, booking or request.

PILOT BASELINE — [workflow], period [start]–[end], recorded by [name]

Columns (one row per item):
date | channel (WhatsApp/call/web) | time received | time first replied | replied by (person/agent) | handled fully by agent? (Y/N) | outcome (booked/qualified/documents complete/no reply) | escalated? (Y/N, reason) | minutes of staff time | notes

Numbers to compute at the end of each period:
- Median first-reply time, and median after hours ([19:00]–[09:00])
- Share replied within 5 minutes; share never replied
- Staff minutes per item, and total hours
- Outcome rate (e.g. enquiries → bookings; requests → documents complete)
- Escalation rate and the top three escalation reasons
- No-show / drop-off rate where relevant

Decision rule agreed before the pilot: continue if [metric] improves by at least [x] with escalation rate under [y]% and no unapproved actions.

Agree the decision rule before the pilot starts. A pilot without one is judged on feelings, and feelings are what the vendor is best at.

The approved facts list

The single document the agent answers routine questions from. If it is not on this list, the agent says "let me check".

APPROVED FACTS — [Business name], owner: [name], last reviewed: [date]

HOURS: [days and times, holidays]
LOCATION: [address, landmark, map link, parking note]
WHAT WE DO / DO NOT DO: [services or products; the common asks you decline]
PRICES THE AGENT MAY QUOTE: [item: price] (anything else: "the team will confirm")
POLICIES: [cancellation, rescheduling, refunds, payment methods, ID requirements]
PEOPLE: [names and roles the agent may mention; which doctor/advocate/teacher does what]
LANGUAGES: [list]
STANDARD ANSWERS: [the ten questions you get most, each answered in two lines]
NEVER SAY: [outcome promises, comparisons with competitors, anything about people's communities or safety, medical/legal/financial opinions]

Review it monthly. Every wrong answer an agent gives in its first month traces back to a stale or missing line here.

The weekly agent review (ten minutes)

Every week the agent is live, by the owner or manager, from the log.

WEEKLY AGENT REVIEW — week of [date], reviewer: [name]

1. Volume: [n] conversations, [n] handled fully, [n] escalated (top reasons: ___)
2. Read five random transcripts. Anything wrong, awkward, or that a customer could misread? ___
3. Read every escalation. Were any avoidable with a new approved fact? Add it: ___
4. Any customer who asked to stop? Confirmed they were stopped: Y/N
5. Running cost this week: [amount]; anything unusual? ___
6. One rulebook line to promote (A → X) or demote (X → A) this week, and why: ___
7. Decision metric this week vs baseline: ___

Cite this page: SigmaticAI, "The prompt and template library for your first AI agent", appsigmaticai.com/library/prompt-and-template-library, updated 1 November 2026.

Link to it rather than copying the table; the numbers are refreshed here and the sources are one click away.