The Lead Triage Desk: sort every new enquiry and draft replies you approve

What you will build

A Lead Triage Desk: one named, repeatable job that reads your new enquiries, sorts them into a few categories you define, drafts replies for the safe ones and flags the risky ones for you. It never sends anything. You read, edit if needed, and press send yourself.

By the end you will have a one-page Desk spec, a tested prompt, an anonymised sample with tricky messages, and a dated known-good copy.

Before you start

• An AI chat you already use: Claude, ChatGPT or Gemini. You do not need to connect any account for this guide.
• 10 to 20 recent enquiries copied from WhatsApp Business, your website form, Facebook or Instagram messages, and email.
• Your price list and service area, even if they are rough. The resource Put your price list in a shape AI can read correctly helps you tidy it first.
• One document to hold the Desk: a Google Doc, Word file or notes page. Call it "Lead Triage Desk".
• About 30 to 45 minutes for the first build.

How it works

A Desk is one repeatable job with written rules and a human owner. Instead of typing "reply to this" and hoping the AI guesses your prices and tone, you write down once what each category means, what a draft may contain, and when the AI must stop and hand the message to you. The AI prepares. You approve and send.

The Desk has three rule sets. Triage rules say how to sort. Draft rules say what a reply may contain. Stop rules say when there is no normal reply, only an "owner review" flag. When a draft is wrong, you fix the rule, re-run the sample, and save the new known-good copy.

1. Map every place enquiries arrive.
2. Choose categories you will actually act on.
3. Write the triage rules in your own words.
4. Write the draft rules.
5. Write the stop rules.
6. Build an anonymised sample with traps in it.
7. Run the Desk prompt.
8. Review like the owner and fix rules, not messages.
9. Save the known-good copy and set a daily routine.

Step 1: Map where enquiries arrive

A Desk can only sort what it sees. If some enquiries sit in a staff member's personal WhatsApp, the Desk gives you a false picture. Start with a plain list of every door a customer can knock on.

ENQUIRY SOURCES
WhatsApp Business: [which number or phone]
Website form: [which inbox it lands in]
Facebook / Instagram messages: [who checks]
Email: [address]
Walk-ins and phone calls: [where someone writes them down]
Staff chats: [any group where leads get forwarded]
Who checks each source, and how often: [name or role]

For now, messages reach the Desk by copy and paste. That is deliberate. You learn the rules on a small, safe sample before you think about connecting an inbox.

Check

• Every source is listed, with the person who checks it.
• You know which two or three sources bring most of your enquiries, from your own experience.
• You have agreed that, for now, enquiries reach the Desk only by copy and paste.

Step 2: Choose categories you will actually act on

Start with five categories. Delete any you will not use, and do not add more until you miss one.

• Hot: inside your service area, needed soon, and enough detail to reply properly.
• Warm: a real job, but missing details you need before you can quote or book.
• Later: a real enquiry, but the timing is months away or they are only comparing.
• Unclear: not enough to judge what they want, where, or when.
• Existing quote: you already sent a quote and the customer has not decided.

"Owner review" is not a category. It is a flag on any message that hits a stop rule. A hot lead asking for a discount is still hot, and also owner review.

Test your list with one question: for each category, what do you do next? If two categories lead to the same action, merge them.

Check

• Each category has a different next action.
• You have five categories or fewer.
• "Owner review" is a flag, not a bucket that hides messages.

Step 3: Write the triage rules in your own words

The triage rules are the heart of the Desk. Generic words like "high intent" mean nothing to the AI. Your words do: "inside Seksyen 1 to 25", "for this month", "gave the unit type". Anthropic's prompting best practices compare the model to a brilliant new employee who lacks context on your norms. Write the rules as the briefing you would give that new staff member.

LEAD TRIAGE DESK
Business type:
Owner (approves every reply before it is sent):
Service area:
Languages we reply in: BM / English / Chinese / Tamil (delete as needed)

TRIAGE RULES
Hot means:
Warm means:
Later means:
Unclear means:
Existing quote means:

FACTS TO LOOK FOR IN EVERY MESSAGE
- Location (area or postcode only, no full address)
- Service or product asked for
- Quantity or size
- Timing or urgency
- Budget or price question
- Language the customer used

"Hot means inside our service area, needed within two weeks, and they told us the service type" is usable. "Hot means serious customer" is not.

Check

• A stranger could sort a message using only your rules.
• Every rule uses words from your business, not marketing words.

Step 4: Write the draft rules

Draft rules decide what a reply may contain, and what the AI is not allowed to invent.

DRAFT RULES
- Reply in the customer's language. If they mix BM and English, you may mix too.
- Sound like a real person at a small Malaysian business: friendly, direct, short.
- Use only prices, services, areas and hours from OUR NOTES. If something is not in our notes, say we will check and confirm.
- Never promise a date, time slot, discount, stock level or result that is not in our notes.
- Ask for at most two missing facts: the ones we need for the next decision.
- Keep it short enough to read on a phone without scrolling.
- End with one clear next step for the customer.
- Mark every draft "WAITING FOR OWNER APPROVAL".

Add one tone line that is truly yours, such as "No emojis unless the customer used them first".

Check

• The rules say what the AI may use, not only what it must avoid.
• There is an explicit "we will check and confirm" fallback for missing prices.
• Every draft is marked as waiting for approval.

Step 5: Write the stop rules

Stop rules are what make the Desk safe to use on real customers. When a stop rule is hit, the AI does not try harder. It stops, says which rule it hit, and hands the message to the owner.

STOP RULES (flag OWNER REVIEW, do not draft a normal reply)
- Complaint, anger, refund, legal threat, medical issue or money dispute.
- The customer says no, later, already booked, wrong number, stop messaging or do not contact.
- The customer asks for a discount, special approval or a date promise.
- The next reply would quote a price or make a promise that is not in our notes.
- The message contains private details (IC, bank, health, family situation).
- You are not confident about the category.

When a stop rule applies: write OWNER REVIEW, the rule it hit, and what decision the owner must make. You may suggest a one-line holding reply only if it promises nothing.

A holding reply like "Terima kasih, saya check dengan boss dulu dan update hari ini" buys time without committing you to anything.

Check

• "Do not contact" and "wrong number" always stop the Desk.
• Discount and date requests always reach you.
• The AI is allowed to say it is not sure.

Step 6: Build an anonymised sample with traps in it

Now collect 10 to 20 real enquiries and clean them. Replace names with Customer A, B, C. Replace phone numbers, emails, full addresses, IC numbers and order numbers. Keep the area if it matters for your rules (for example "Taman Molek" instead of the full address).

Make sure the sample covers every category, plus traps: a discount request with a same-day deadline, "stop messaging me", a complaint, a message in BM, an enquiry outside your area, and a bare "hi, how much?". A Desk that only sees perfect messages fails on the first messy one.

SAMPLE ENQUIRIES (anonymised)
#1 | WhatsApp | Mon 9:10am | "..."
#2 | Website form | Mon 11:40am | "..."
#3 | Email | Tue 8:05am | "..."

Save this sample in your Desk document. You will reuse it every time you change a rule.

Check

• No real name, number or address is left in the sample.
• Every category and every trap appears at least once.
• The sample is saved, not only pasted into a chat.

Step 7: Run the Desk prompt

Open a fresh chat. Paste the prompt below with your filled spec, your notes and the sample. Anthropic's guide to reducing hallucinations recommends allowing the model to say it does not know, and restricting it to the documents you provide. Both are built into this prompt.

You are my Lead Triage Desk. You prepare work for me. You never send messages, contact anyone or decide prices.

Use ONLY the Desk spec and our notes below. If something is not in them, say so. If you are unsure about a category, choose Unclear and explain why.

For each enquiry, return:
1. Enquiry number and source
2. Category (from my list only)
3. One-line reason, quoting the words in the message that led to it
4. Missing facts I still need
5. Draft reply that follows my draft rules, OR "OWNER REVIEW" with the stop rule it hit and the decision I must make
6. Status: WAITING FOR OWNER APPROVAL

After the list, give me a short summary: how many in each category, and which messages need me first.

DESK SPEC:
[paste triage rules, draft rules and stop rules]

OUR NOTES:
[paste price list, service area, hours, policies]

ENQUIRIES:
[paste anonymised sample]

Asking for the words behind each category gives you something to check. No words, no category: it is a guess.

Check

• Every enquiry appears once, with nothing merged or skipped.
• Every draft ends with "waiting for owner approval".
• Every trap message is flagged OWNER REVIEW, not answered.

Step 8: Review like the owner and fix rules, not messages

Read the output as if you were about to press send on every draft. Mark each one with one of three states:

• Verified: you checked it against your notes and would send it as is.
• Unverified: it looks fine, but you have not checked a fact yet.
• Broken: it invented something, missed a stop rule, or put a lead in the wrong category.

For every broken item, ask which rule was missing or unclear, and add one line to the spec. Then run the whole sample again, not only the broken message. A new rule can fix one case and break another.

REVIEW NOTE
Date:
Desk version:
Broken items: [number, what went wrong]
Rule changed: [the exact line added or edited]
Result after re-run: [all verified / still broken: ...]

Check

• No draft contains a price, date or promise missing from your notes.
• Every broken item led to a rule change, not only a hand edit.
• The re-run passed on the full sample.

Step 9: Save the known-good copy and set a daily routine

When the full sample passes, save a known-good copy with today's date: spec, prompt, notes, sample and passing output. If a later edit makes the Desk worse, go back to this copy instead of guessing what changed.

Then set a simple routine:

DAILY ROUTINE (morning and after lunch)
1. Copy new enquiries, anonymise them.
2. Run the Desk prompt in a fresh chat or your Desk project.
3. Handle OWNER REVIEW items first.
4. Read, edit and send the safe drafts yourself.
FRIDAY: review broken drafts, change a rule if a mistake repeated, save a new known-good copy.

WhatsApp Business lets you label chats (for example "New lead" or "Needs follow-up"). Menus may move; check your app, or see How to use labels in the WhatsApp Help Center.

Check

• A dated known-good copy exists in your files, not only in a chat.
• The routine names who runs the Desk and who approves.
• There is a weekly slot to improve rules.

Worked example

This example is made up for teaching. The business, messages and details are not real.

A tuition centre in Johor Bahru teaches primary and secondary students. Parents write in English, BM and Mandarin. The owner approves every reply. The notes say: classes run in Taman Molek only, subjects are Maths, Science and English for Standard 4 to Form 5, and only the owner confirms class slots.

Part of the Desk spec:

Hot means: child's level and subject given, parent is in JB, wants to start this month.
Warm means: real interest but level or subject missing.
Later means: asking for next year or "just checking fees".
Unclear means: no level, no subject, no timing.
Stop rule added after first run: never name a start date or class slot. Say the owner will confirm the slot.

Three sample messages, anonymised:

#1 | WhatsApp | "Hi, anak saya Form 2, nak ambil Maths dengan Science. Rumah dekat Taman Molek. Boleh start bulan ni?"
#2 | Website form | "How much for tuition?"
#3 | WhatsApp | "Last month you charge me double. I want refund, if not I report."

What the Desk returned on the first run, in short:

• #1 Hot. Reason: "Form 2", "Maths dengan Science", "Taman Molek", "bulan ni". The BM draft also said "boleh mula minggu depan", a promise not in the notes. Broken.
• #2 Unclear. Missing: child's level, subject, area. Draft asked for level and subject only, in English, and ended with a clear next step.
• #3 OWNER REVIEW. Stop rules hit: complaint, refund, money dispute. No normal reply. Suggested holding line: "Thank you for telling us. The owner will check your record and reply to you today."

The fix: the owner added the stop rule shown above ("never name a start date or class slot"), ran all the sample messages again, and #1 came back with: "Terima kasih! Untuk Form 2 Maths dan Science, kami ada kelas di Taman Molek. Anak lebih suka hari biasa atau hujung minggu? Saya akan sahkan slot kelas selepas itu." Verified, sent by the owner. The spec, sample and output were saved as the new known-good copy.

Common mistakes and fixes

• The AI invents prices → add "Use only prices from our notes; otherwise say we will confirm", and check your price list is actually pasted in.
• Every lead comes back "Hot" → your rules are too loose. Add the facts a hot lead must include, such as area plus timing plus service.
• A "stop messaging me" reply gets a cheerful follow-up draft → add "do not contact" and "wrong number" to the stop rules and put that trap in your sample.
• You fix drafts by hand every day → change the rule once and re-run the sample.
• The Desk works in one chat and not the next → the rules only lived in that chat. Keep the spec in a file, or put it in a Project.
• Real names end up in the AI tool → anonymise before pasting, every time. Make it step one of the routine.

Take it further

• Make it a Project. Put the Desk spec, price list and FAQ into a Claude Project (or a ChatGPT Project or Gemini Gem) so you do not paste them every time. See Set up a Claude Project as your business brain.
• Make it a Skill. Once the Desk is stable for a few weeks, package it as a Claude Skill so Claude applies it whenever you paste enquiries. See Write your first Claude Skill for quotes or enquiry replies.
• Make it a weekly habit. Every Friday, review broken drafts and update one rule. The Desk improves because the rules improve.
• Related resources: Draft WhatsApp replies in BM or English that you check and send, Follow-ups that stop when they should, Check the AI's answer before a customer sees it and Before you paste anything into AI: a two-minute safety check.

Quick checklist

• All enquiry sources listed, with who checks each one.
• Five categories or fewer, each with a different next action.
• Triage, draft and stop rules written in your own words.
• Stop rules cover complaints, refunds, discounts, date promises and "do not contact".
• Every draft marked "waiting for owner approval".
• Broken drafts lead to rule changes, then a full re-run.
• Known-good copy saved with a date.
• A person reads and sends every reply.