how to build an automated sales machine (FULL GUIDE)

this guide builds something that finds customers, sells to them and keeps them paying, without you doing any of it once it’s set up.

i’m using web services as the example because it’s the easiest one to start with, but the same setup works for whatever you’re selling.

this space is saturated but there’s still a lot of money in it if you go about it differently.

you narrow a huge list down to the few businesses that genuinely need the service, you build it for them before you make contact, and you reach them.

this guide builds that as one system with claude code, and we wrap it in an api at the end, and setup a hermes agent runs it on a schedule.

here’s what it does once it’s live:

• finds businesses worth contacting and filters out everything else

• builds the site before anyone has spoken to them

• reaches the owner through their own contact form or by post

• takes payment and switches the site live

• runs a second offer on top that pays every month

before you build anything, pick the niche

what you want is a business where the customer researches before they buy, where the job is worth a few thousand at least, and where turning up without a website costs them work.

solar, roofing, hvac, pools, driveways, extensions, landscaping, windows, flooring, garage doors, fencing, tree surgery.

these businesses are also used to spending on getting customers.

pick ten and only watch those. a narrow list means your templates fit properly, your outreach is specific, and the running cost stays close to nothing.

which model does what

every step in this uses a different model.

claude opus 5 orchestrates. it’s the one hermes runs on and the one deciding what happens next.

gemini 2.5 flash reads the listings and the sites. it’s cheap and it handles pages properly, which matters when you’re putting a few thousand through it.

kimi k3 builds the templates. it’s the best of them at frontend, so the sites come out looking really good

gpt image 2 does the images for the website, postcard fronts and the review renders.

claude haiku reads the client emails and updates the sites. it’s a small job and it runs constantly, so it doesn’t need anything bigger.

step 1, the finder

open claude code in an empty folder and start here.

“build me a node project that finds local businesses in a given niche and area.

two data sources:

first, new business registrations. before you write anything, search for the public sources that list new registrations, trade licences or permits in [region], tell me what each one exposes, how often it updates, and whether there’s an api or i’d need to scrape it. then use the best one. if there isn’t a usable source, say so instead of building around one that doesn’t work.

second, google maps. use a scraper on apify with my apify api key to find a maps scraper and pull every business in the niche for a given area, with name, address, website, phone, rating and review count. (you can use the official google places api for this too)

output both into one sqlite table with a source column so i can tell where each record came from.

give me a single function findBusinesses(niche, area) that runs both and returns the combined list.“

step 2, the filters

next prompt:

“add a scoring pass to that project.

for each business, use gemini 2.5 flash to look at their listing and their site if they have one, and return:

• whether they have a site, and if so whether it works on a phone or breaks

• when it looks like it was last touched

• whether the hours, phone and address are on the page

• whether it looks owner run or chain owned

• when they last posted or replied to a review

• whether there’s a working contact form and what fields it has

then score each business out of 10 as a prospect and give me the single worst thing about their current setup in one line.

drop anything already on a modern platform with a decent site, anything chain owned, and anything with no activity in the last year.

if a page won’t load, mark it as unknown rather than scoring it anyway.

write the scores back to the table and give me a function scoreBatch(limit).“

what comes out is a ranked list, the specific problem you’ll mention when you contact each one, and which channel they go into.

step 3, the templates

don’t generate a site from scratch each time. build ten and swap the data in.

use kimi k3 via openrouter/replicate for this one,

“create a templates folder with one static site template per niche.

each template is html, tailwind and a single data.json. every piece of copy, every service, every price and every image path reads from that json.

the solar template needs a quote form, service area, finance options, a gallery of past jobs and a savings section. the roofing one needs emergency callout, a gallery, service area and a quote form. build the rest to match what that trade’s customers actually look for.

then write buildSite(businessId, templateName) that pulls the business record from the database, maps it into data.json, renders the template into a folder, and returns the local path.“

that turns the build into a data merge. the cost per site drops significantly and the output stays consistent.

step 4, the outreach

two channels, decided by whether they had a contact form.

“add an outreach module.

for businesses with a contact form, write fillForm(businessId) that opens their site with playwright, finds the form, and submits one line saying who i am, the specific problem from their score, and the link to their built site. rate limit it to a handful a day with random gaps.

for businesses without one, write makePostcard(businessId). take a screenshot of their built site, pass it to gpt image 2 with the business name to produce the postcard front, and leave the bottom right corner empty.

then generate a real qr pointing to their site url with a tag on the end, composite it into that empty corner, and output a print ready pdf.

then sendPostcard(businessId) which posts it through the lob or postgrid api using the address from the record.“

owners usually ignore email because they get hundreds of them. they read contact form submissions because they assume it’s a customer. keep the volume low and make every one specific.

step 5, the paywall and the payment

the qr goes to the finished site, not a page about the finished site.

“add a paywall and checkout.

every built site deploys to a subdomain with a bar across the top saying it goes live when they subscribe. everything else on the page is visible and working.

use whop for checkout. create the product, put the checkout link in that bar, and listen for the payment webhook.

on payment: remove the bar, point their domain if they have one, mark them paid in the database, and send a welcome email with the address they can email changes to.“

they see their own business working properly on their phone. they either want it or they don’t, and the ones who want it are already looking at it.

step 6, the edit agent

this is what makes the subscription worth paying in month two.

these owners aren’t logging into a dashboard. a roofer isn’t technical so he just emails you. new price on gutter cleaning, we cover this town now, here’s a photo from friday’s job.

“add an inbox agent running on claude haiku.

watch a shared mailbox. when an email comes in, match the sender to a client in the database.

read what they’ve asked for, work out which fields in their data.json it affects, make the change, rebuild the site and deploy it.

attached photos go into their gallery folder and get added to the gallery array.

reply to the email confirming what changed.

if the request is ambiguous or would change pricing, reply asking one question instead of guessing.“

step 7, the review module

reviews are the second offer and the one they’ll actually notice working.

“add a review module.

first, a finder. compare each business’s review count against the average for that category within a few miles, and flag anyone significantly below it. sort by the size of the gap.

then onboarding for anyone who buys it. they upload one photo of the owner, a van or a finished job. generate one image from it with gpt image 2, with a blank card in frame. store it as their base image.

then a sender. when a new customer arrives from their csv or booking system, composite that customer’s name onto the blank card in the base image and email it with a short thank you and a review link.

one image generation per client. every customer after that is a text overlay on the cached image.“

on the rules, everyone gets asked, not just the ones you think were happy. that line is what separates a review system from review gating, and gating is what a lot of reputation software is quietly doing.

platforms have started enforcing on it and there are real fines attached in some places. it goes out after the job is finished.

text messaging is the upgrade. sending business messages from a number needs carrier registration in most places, and under the business’s own name they need their own rather than yours.

step 8, wrap it in an api

everything so far is functions. this is what lets something else drive them.

“wrap every function in an express api with token auth.

POST /find takes a niche and area and runs the finder. POST /score runs the scoring pass on unscored records. POST /build builds sites for everything above a score threshold. POST /outreach fills forms and queues postcards. POST /send posts the queued postcards. GET /pending returns anything waiting for my approval. POST /approve takes ids and releases them. GET /status returns counts by stage.

deploy it to railway and give me the base url and the token.“

step 9, hand it to hermes

“i want a hermes agent (https://nousresearch.com) running on claude opus 5 that calls this api on a schedule. here’s the base url and token, and here’s the openapi spec.

set it up with a telegram bot so it can message me, and host it on railway.

daily: check the registration feeds, build sites for anything new in my niches, and put them in pending. weekly: scrape one new area and score everything in it. daily: fill contact forms for anything approved, low volume with gaps between them. weekly: queue postcards for approved businesses without forms. daily: check for new payments and confirm the sites went live. daily: for review clients, send images to any new customers.

message me each morning with what’s pending and wait for approval before anything goes out.“

then it pings you on telegram with the batch ready. you look at the sites, cut anything that came out wrong, approve, and the cards go to print.

what to charge

$200 a month for the site. just five clients is $1,000/month.

$99 a month for reviews up to a few hundred customers, $199 above that.

or bundle both at $250, which raises what the average client pays and makes them much harder to lose.

say the edit thing on the paywall. email us anything and the changes go live the same day.

what it costs to run

the templates mean building a site is a data merge rather than a generation, and hosting a static site is close to nothing on top.

the review images are one generation per client and then a text overlay, so effectively zero per customer.

postcards are the only real cost at a few dollars each including postage, and those only go to businesses that scored well.

so a client paying $250 a month costs a few dollars a month to serve. that’s what makes twenty clients a real business rather than a busy one.

everyone else is emailing and cold calling businesses. you’re turning up with the product already built, to the people who actually needed it.

join here for more value: https://t.me/+pbCBBtUEtu1lZDA1