The prompts from the video
The AI lead generation system
These are the prompts behind each step of the lead system in my video, from finding the right people to the first reply and the weekly numbers. Paste them into ChatGPT or Claude and fill in the blanks. Start with A1: it tells you which link to fix first.
The seven links
- 1The offerA stranger can repeat what you sell and the result in one sentence.A2A3
- 2TrafficYou can turn up a source of new prospects on purpose this week.A4A7A8
- 3The pageThe page they land on screens out bad fits.
- 4Speed to leadA new enquiry gets a reply inside 5 minutes.A5A6
- 5Follow-upSomeone follows up at least three times if they go quiet.A6A11
- 6The CRMEvery lead is in one place with its source and a named owner.A5A11
- 7TrackingYou can say which source produced last month’s customers.A9A10
The prompts
How to use them
Fill in every
[blank in square brackets]. Delete any line that doesn’t fit your business.Paste the prompt into ChatGPT, Claude or any AI chat as a new conversation.
The AI must never invent a number, a fact or a result. Every prompt tells it to write UNKNOWN instead. If it invents one anyway, delete it.
Examples are examples. Nothing on this page is a result anyone got.
Check the output before you use it. Each prompt ends with how.
A1Find your weakest link (the seven links)
What it does. It asks one yes-or-no check per link of your lead system and names the weakest link and one fix you can make this week. It is a quick check, not the full audit: the full seven-layer score with the 90-day route is the paid Lead Generation GPS.
When to use it. Before you buy any tool, hire anyone or add any AI. Run it again after each fix.
You are doing a quick check of a service business's lead system. Do not invent facts. If I haven't told you something, mark it UNKNOWN. My business: [what you sell, to whom, in which area] Answer one question per link with YES, NO or UNKNOWN, using only what I tell you: 1. The offer: could a stranger repeat what I sell and the result in one sentence? 2. Traffic: can I turn up a source of new prospects on purpose this week? 3. The page: does the page they land on screen out bad fits? 4. Speed to lead: does a new enquiry get a reply inside 5 minutes? 5. Follow-up: does someone follow up at least three times if they go quiet? 6. The CRM: is every lead in one place with its source and a named owner? 7. Tracking: can I say which source produced last month's customers? Here is what I have: [a few lines of notes] Return: - The seven answers in a short list. - The weakest link: the first NO in the order above, and one sentence on why it limits the links after it. - One fix for that link I can make this week, and how I'll know it worked (a number and a date). Do not score the links, build a plan or recommend tools.
What a good output looks like.
- Seven short answers, UNKNOWN wherever you gave no evidence.
- One named weakest link, with a reason that refers to your notes.
- One fix with a check and a date.
It will often mark tracking UNKNOWN. That is usually the honest answer.
Check it.
- Every YES should point to something you wrote.
- A YES with no evidence is the AI being polite. Change it to UNKNOWN.
A2Write the one-sentence offer
What it does. It drafts your offer in a fixed format and tests each version. It writes three pain statements in your customers' words, and it flags every claim you can't back.
When to use it. Before any ad, page or script. Every other piece repeats this sentence.
Help me write the one-sentence offer for my business. Use this format:
We help [who] get [specific result] in [timeframe] without [the thing they hate].
Add a risk reversal only if I give you one I can honour.
About my business:
- What I sell: [service]
- Who buys it (trade, size, area): [who]
- The result customers actually get, in their words: [result, or UNKNOWN]
- How long it takes: [timeframe, or UNKNOWN]
- What they hate about the alternatives: [the thing]
- Proof I hold right now (numbers, reviews, case studies I can show): [list, or NONE]
Then:
1. Write five versions.
2. Score each against four tests: a stranger in my market gets it on first
hearing; the result is specific ("more leads" is not a result); it names one
buyer; every claim is backed by proof I listed.
3. Write three pain statements in my customers' words: first person, present
tense, something a real customer could say on a phone call. Not summaries
like "unreliable lead quality".
4. List every claim in your versions that needs proof I don't have, and rewrite
those versions without it.
Never add a guarantee, a number or a result I didn't give you.What a good output looks like. Five sentences, a scoring table, three pain statements that sound spoken, and a list of claims that need proof.
Check it. Search the output for any number or guarantee you didn't type. Delete it.
A3What you can pay for a booked call
What it does. It works backwards from what a customer is worth to the most you can pay per customer, per showed call, per booked call and per lead. It shows every step, gives worst and best cases, and writes down what "working" means before you spend.
When to use it. Before the first ad. It sets your budget, your kill rule and your scale ceiling.
Work out the most I can pay for a customer, a showed call, a booked call and a lead. Show every step. Multiply by the rates. Never divide by them. My numbers (I've written ESTIMATE next to any I'm guessing): - Revenue from one new customer: [$] - Gross margin (the share left after delivery costs): [%] - Share of that gross profit I'm willing to spend to win one customer: [%; a third is a common starting point] - Close rate (showed calls that become customers): [%] - Show rate (booked calls that actually happen): [%] - Booking rate (leads that book a call): [%] Steps: 1. Allowable cost per customer = revenue x gross margin x share. 2. Allowable cost per showed call = step 1 x close rate. 3. Allowable cost per booked call = step 2 x show rate. 4. Allowable cost per lead = step 3 x booking rate. Then give a worst, expected and best case by moving each rate I marked ESTIMATE down and up by a quarter of its value. Then write two sentences: what "working" means (a cost per booked call at or under step 3, at a stated number of leads a week, by a stated day) and what "kill it" means. If the allowable cost per lead is lower than my market can plausibly deliver, say so plainly: that is an offer or sales problem, not an ad problem.
What a good output looks like. Four numbers, each smaller than the one before, with the arithmetic shown. A worst, expected and best table. Two plain sentences.
EXAMPLE, to check the AI's arithmetic. These are not anyone's numbers:
| Step | Arithmetic | Result |
|---|---|---|
| Per customer | $9,000 × 50% × one third | $1,500 |
| Per showed call | $1,500 × 25% close rate | $375 |
| Per booked call | $375 × 80% show rate | $300 |
| Per lead | $300 × 30% booking rate | $90 |
Check it. If any number is bigger than the one above it, the AI divided. Run it again.
A4Build a lead list from public records
What it does. It finds the official public records that list the businesses you sell to. It writes the filter and the cleaning rules, sets up the columns you fill in by hand, and lists the contact rules.
When to use it. When you sell to a type of business that is licensed or registered: insurance agencies, law firms, HVAC and other licensed trades.
I sell [service] to [type of business, e.g. insurance agencies / law firms / HVAC companies] in [state or area]. Help me build a list of every one of them from public, official records. 1. List the official public sources that publish these businesses: licensing boards, state bar or insurance department lookups, open-data portals, secretary of state records. For each one give: the name, the exact page, who publishes it, the fields it includes (name, licence number, status, address, phone, email, website, licence date), whether a bulk download exists, how often it updates, and its terms of use. If you are not sure a source exists, write NOT VERIFIED. Do not invent web addresses. 2. Write the filter: which status values to keep (active only), which licence types match my buyer, and the area filter. 3. Write the cleaning rules: duplicates (same phone or same licence number), missing contact details, addresses outside my area, businesses that look closed. 4. Add three columns I will fill in myself: website; "already advertises" (from the Meta Ad Library, with the start date of their oldest active ad); and one fact I checked, for the first message. 5. List the rules for contacting businesses from this list in my country: consent for texts and calls, email rules, opt-outs. Mark it as general guidance, not legal advice. Do not give me a count. I will get the count by downloading the file.
What a good output looks like.
- A table of three to eight sources, some marked NOT VERIFIED.
- A filter and a cleaning checklist.
- Three empty columns.
- A short contact-rules list.
Check it.
- Open every web address. Delete any that doesn't load or isn't official.
- Download the file, apply the filter, and note the count and the date yourself.
A5Research a lead before you reply
What it does. It turns a new enquiry into a one-page brief on their business before anyone replies: who they are, how they get customers now, their likely weak link, which of your options fits, and what to ask. Every fact carries its source.
When to use it. Every time a lead comes in, before the first reply. Use it again before a booked call.
A new lead just came in. Before anyone replies, build a one-page brief on their business. What they gave us: [name, business name, website, the message they sent, form answers, where they came from] Our options, so you can say which fits: [list your offers, one line each, no prices] Use only public business information: their website, their Google Business Profile and reviews, their company LinkedIn page, the Meta Ad Library, news. Give the source link for every fact. Label anything you infer as INFERENCE. If you can't find something, write UNKNOWN. Never look for or guess personal phone numbers, personal emails, or anything about their private life. Return: 1. Who they are: what they sell, to whom, where, and size signals (team size, locations, review count), each with its source. 2. How they get customers today: ads running (with start dates), website forms, booking links, review volume. Evidence only. 3. The likely weak link in their lead system, as a hypothesis: the evidence for it, and what would prove it wrong. 4. What they asked for, in their own words. 5. Which of our options fits, why, and which one doesn't fit. 6. Three questions to ask in the first reply or on the call. 7. What not to say: claims we can't back, and assumptions the evidence doesn't support.
What a good output looks like.
- A one-page brief where most lines carry a link.
- One to three INFERENCE labels, and UNKNOWN where the public record is thin.
- The weak link is written as a hypothesis, not a verdict.
- Three specific questions.
Check it.
- Click three of the sources.
- Strike any fact without a link.
- If a personal phone number or personal email appears, delete it.
A6Answer a new lead by text
What it does. It tells an AI how to answer a new lead by text the way you would.
- It asks one question at a time.
- It never quotes a price.
- It offers two real times.
- It hands off to a named person when it should.
Run the ten tests before it texts anyone.
When to use it. For leads that come in through a form and agreed to be texted. It works in a chat for testing, then in the AI step of your automation tool.
How to use it.
- Fill in every [blank]. The two open times can be any two real slots for now.
- Open a new chat. Paste everything from COPY FROM HERE to COPY TO HERE as your first message. It should reply READY.
- Send each test message below, one at a time. Mark each one passed or not.
- Every fail becomes a new rule at the end of Part 2. Then start a new chat and run all ten again.
- When all ten pass, connect it to your form, CRM and texting number. First delete the last paragraph, the one starting TESTING IN A CHAT, or it may text READY to a lead.
- Keep a person on price questions and complaints. Rules 2 and 7 hand those off; don't remove them.
- Set your automation to act on the two words the AI writes on its own, and to strip both before anything is sent:
- HANDOFF: send the message above it, pause the automatic texts and alert your named person.
- NO_REPLY: send nothing.
Text only people who agreed to be texted on your form. That is general guidance, not legal advice. Settle the wording of that consent box with your own adviser.
============ COPY FROM HERE ============ PART 1: THE CONTEXT You answer new enquiries by text for [Business], a [kind of business] in [town or area]. We do: [service 1], [service 2], [service 3]. We don't do: [services you don't offer]. We cover these ZIP codes: [service ZIP codes]. Your one goal: get the lead to pick a time for [a booked inspection / a quote visit / a call]. Price questions go to [who takes price questions]. You never answer them yourself. Booking times come from [where booking times come from]. The two open times are [time 1] and [time 2]. Offer only those. [Named person] takes over whenever you hand off. The lead's first name is [first name] and they asked about [service]. When this is connected to your form, your automation fills in those two. Before you offer times, you need two answers: how soon they need it, and the ZIP code where the work is. Ask for whichever is missing, one question at a time. Your first question: "[how soon they need it, as a question]" Your second question: "What's the ZIP code where the work is?" Once you have both, and the ZIP code is one we cover, offer the two open times. PART 2: THE RULES 1. Ask one question at a time. Never two in one message. 2. Never quote a price, a range or an estimate by text. Say that [who takes price questions] prices every job and will reply to them here, then hand off. 3. When they're ready to book, offer the two open times, not a link. Ask them to reply 1 or 2. 4. Use their first name and the job they asked about. 5. If they reply STOP, or say they don't want texts, write NO_REPLY and nothing else. Never text them again. 6. If they ask whether you're a person or a bot, say you're an assistant for [Business], and offer to have [Named person] reply. 7. Hand off when you're unsure, when they're upset or complaining, or when they ask for something these rules don't cover. To hand off, send: "Thanks, [first name]. I've passed this to [Named person], who will reply to you here." Then write HANDOFF on a new line. Also: - Keep every message to two short sentences plus your one question. - Never make up a fact about [Business]. No prices, dates, warranties, discounts or promises that aren't in Part 1. - If they ask who this is, say you're texting from [Business] about the [service] they asked about on our form, then ask your next question. - If the job is outside our ZIP codes, or it's a service we don't do, say so politely and offer no times. - If they say they've already hired someone or don't need it, send: "No problem, thanks for letting us know. We won't text again." Then write NO_REPLY to anything after that. - If they ask you to call them, send: "Sure, [first name]. I've passed this to [Named person], who will call you." Then write HANDOFF on a new line. - If a message starts NO REPLY, the lead hasn't answered. Send this once: "Hi [first name], just checking you saw our text about [service]. Still want help with it? A yes or no is fine. Reply STOP to opt out." If you get NO REPLY a second time, write HANDOFF. PART 3: EXAMPLES The lead writes: "The ZIP is [a ZIP code you cover]. Can you come out today?" A good reply. Copy this style: "Thanks, [first name]. [Named person] can come out [time 1] or [time 2]. Which suits you better? Reply 1 or 2." A bad reply. Never send anything like this: "Hi, yes we can come today. It's usually around $[price]. Book here: [link]. What's your address and what's the problem?" Why the bad one fails: it quotes a price, sends a link instead of two times, asks two questions at once, and says yes before anyone checked the calendar. TESTING IN A CHAT: reply to this first message with READY and nothing else. Treat every message after it as a text from the lead. (Delete this last paragraph before you connect the prompt to texting.) ============ COPY TO HERE ============
The ten tests. Send each one as the lead. Every fail becomes a rule.
| # | Send this | A pass looks like |
|---|---|---|
| 1 | How much will it cost? | Quotes no price, range or estimate. Says the person who prices jobs will reply here, then writes HANDOFF |
| 2 | Can you come today? | Treats it as urgent, asks for the ZIP code if it doesn't have it, then offers the two open times. No link, and no yes to today unless one of the times is today |
| 3 | Who is this? | Says it's texting from your business about the service they asked about, then asks one question |
| 4 | Is this a bot? | Says it's an assistant for your business and offers to have your named person reply |
| 5 | STOP | Writes NO_REPLY and nothing else |
| 6 | We already hired someone. | Sends "No problem, thanks for letting us know. We won't text again." Then nothing more |
| 7 | The job is in [a ZIP code you don't cover]. | Says politely that it's outside your area and offers no times |
| 8 | Do you do [a service you don't offer]? | Says you don't, invents nothing, offers no times; hands off if unsure |
| 9 | Just call me. | Says your named person will call them, then writes HANDOFF |
| 10 | NO REPLY: the lead hasn't answered for a day. | Sends the one follow-up, once. A second NO REPLY gets HANDOFF |
What a good output looks like. READY to the first message, then short replies that follow the rules. Expect one or two fails on the first run. Each fail becomes a rule.
Check it. All ten pass in a fresh chat, after your last rule change.
A7Ten ad angles that filter
What it does. It writes ten ad hooks: two of each of five types. The first line names who the ad is for, and every claim is tied to proof you actually hold. Then it picks three to test first.
When to use it. When you build an ad batch. Use it again every two to four weeks, when ads tire.
Write ten ad hooks for [business], selling [service] to [who] in [area].
Our offer in one sentence: [from A2]
Proof we can show: [list, or NONE]
Pain statements in our customers' words: [from A2]
Rules:
- Two hooks of each of these five types:
1. Question: names a cause they haven't considered.
2. Bold claim: only if it maps to proof listed above. If there is no proof,
write two more Devil's advocate hooks instead.
3. Devil's advocate: agrees with their cynicism first. Needs no proof.
4. Story open: first person, drops straight into a scene.
5. Proof-led: a result plus the artifact that shows it. Only with real
proof; otherwise write two more Question hooks.
- The first line calls out who it's for with a real qualifier (the trade, the
area, the size).
- Each hook must land in one frame. If it needs a second sentence to make
sense, rewrite it.
- No guarantee, price, number or result I didn't give you. Nothing "free" and
nothing "without spending money".
Return a table: type, hook, the qualifier in it, the proof it relies on
(or NONE), passes the one-frame test (yes/no).
Then pick the three to test first, and say why.What a good output looks like.
- A ten-row table.
- Where you have no proof, the bold-claim and proof-led rows are swapped for other types, and the output says so.
- Three picks with reasons.
Check it. Any hook that claims a result you didn't give, delete it. Read each hook alone; if it needs more context, it fails.
A8Turn one article into a week of posts
What it does. It turns one article you wrote into a LinkedIn post, a carousel outline, three short video scripts and a newsletter paragraph, adding nothing that isn't in the article.
When to use it. Each time you publish an article.
Here is an article I wrote: [paste it] Turn it into a week of posts. Add nothing that isn't in it. 1. One LinkedIn text post, under 1,300 characters: a first line that stops the scroll, one idea, one ask. 2. A carousel outline: 6 to 8 slides, one sentence each, the last slide is the ask. 3. Three short video scripts, 40 to 60 seconds each at 2.5 words a second: the first line to camera, the middle over a screen (say what's on the screen), the ask to camera. 4. One paragraph for my newsletter. Rules: each piece makes one point from the article. No facts, numbers or claims that aren't in the article. The ask is the same everywhere: [your ask, e.g. a word to comment, or a link].
What a good output looks like. Six pieces, each with one point you recognise from your article, all ending on the same ask.
Check it. Any sentence you can't find a source for in the article, delete it.
A9Kill, keep or raise an ad
What it does. It applies five written rules to your ad numbers and returns one decision per ad, with the rule and the numbers behind it. "Not enough evidence" is an allowed answer.
When to use it. Daily while ads run. Write the rules before you read any result.
Decide what to do with each ad using only these rules. Don't change the rules, and don't use judgement outside them. My targets (from A3): target cost per lead [$], target cost per booked call [$], planned daily spend [$]. My rules: 1. Kill: spend at least 3x target cost per lead with zero results -> pause it. 2. Underperformer: cost per result at least 1.5x target over 7 days, on 5 or more results -> pause it and review before restarting. 3. Fatigue: frequency at least 3.5, or click-through down 40% against its own 14-day average -> flag only. Refresh the creative; don't pause it. 4. Scale: at or under target for 3 or more days in a row -> flag it for a person to raise the budget. Never raise a budget automatically. 5. Circuit breaker: account spend today at least 1.5x planned -> pause everything and alert me. The rows: [paste: ad name, days live, spend, results, cost per result, frequency, click-through, its own 14-day average click-through] For each ad return: the decision (kill / pause and review / flag fatigue / flag to scale / keep / NOT ENOUGH EVIDENCE), the rule that triggered it, and the numbers behind it. If no rule triggers, the answer is keep. If there is too little data to apply a rule, say NOT ENOUGH EVIDENCE. Never call an ad a winner.
What a good output looks like. One row per ad. Most young ads are "keep" or NOT ENOUGH EVIDENCE. Every decision names its rule and quotes the numbers.
Check it. Recompute one decision by hand. If any budget was raised without a person, the setup is wrong.
A10The weekly numbers
What it does. It turns a week of numbers by source into cost per lead, per booked call and per customer, and the rates between them. It names the one stage holding everything back and sets one action for next week.
When to use it. Every Monday.
Here are last week's numbers by source: [paste rows: source, spend, leads, booked calls, showed calls, customers, revenue] My allowable costs (from A3): [paste] Return: 1. A table per source: cost per lead, cost per booked call, cost per customer, and the rates (lead to booked, booked to showed, showed to customer). Leave a cell blank when its count is zero: zero customers is not a zero cost. 2. The stage that is furthest below where it needs to be, against my allowable costs. That is the constraint. 3. One action for next week aimed at the constraint, and the number that will tell me if it worked. Label every number MEASURED if it came from my rows. Don't forecast unless I ask. If you do, label it MODELLED and state the assumptions.
What a good output looks like. A small table with some blanks, one named constraint, one action and one number to watch.
Check it. Recompute one cost: spend divided by that count. This is the one place division is right, because you are measuring an actual cost.
A11Turn a sales call into CRM notes
What it does. It turns a call transcript into a CRM note: their situation, their pains in their words, what was promised, the next step with an owner and a date, and a follow-up draft.
When to use it. Right after every sales call.
Here is the transcript of a sales call: [paste] Write the CRM note. 1. Their situation, in three sentences. 2. Their pains, in their own words (quote them). 3. What they said about budget, timing and who decides. Only what they said; otherwise write NOT DISCUSSED. 4. Objections they raised, and how each was answered. 5. Anything we promised, in the exact words used. 6. The next step, who owns it, and the date. 7. Which of our options fits: [list your options] and why. 8. A follow-up message for me to send, under 120 words, that restates the next step. Don't invent commitments, numbers or dates that aren't in the transcript.
What a good output looks like. A note you could hand to someone else and they'd know exactly where the deal stands. Several NOT DISCUSSED lines are normal.
Check it. Every quote should be findable in the transcript.
Get every prompt as one file
Want it built with you?
Lead Generation Group installs this system inside your business, owned by your team.