HIGHLY EVOLVED PLANT

Small Business Advertising with AI

Building a Knowledge Vault of AI Information, Tools, News & Workflows to keep you informed on the latest in AI for your need

[HIGHLY EVOLVED PLANT @TwelveCacti]

Morning Do

You are writing your website in your own words and selling to people who use different ones. Fix that in six minutes with data already sitting on your phone. Open your texts and your email, pull the last twenty messages customers sent YOU, the inquiries, the questions, the complaints, the ones that came in at nine at night, and paste them in with this: "These are twenty real messages from my customers. One, list every word and phrase they use to describe what I do and what they need. Two, list the words I use on my website that none of them used, here is my homepage text. Three, tell me the three things they worry about that my website never mentions." The first list is your headline. The second list is what you delete. The third one is the interesting one, because it is almost never price. It is usually mess, timing, or whether you will actually show up, and those three fears are why somebody chose the other guy who happened to say the words out loud. You are not guessing at customer language. You have twenty samples of it, and you have had them the whole time.

[HIGHLY EVOLVED PLANT @TwelveCacti]

Here is a quieter way to lose money than any of the obvious ones. You launch forty ads by hand over a couple of evenings, and somewhere around number 37 you fumble the tracking on one and name three of them things like "final_v2_use_this." Two weeks later you are looking at a report where the best performer cannot be traced back to an actual creative, so you scale the one you can identify instead of the one that worked. Nothing broke. No error message. You just spent the next month's budget on your second-best ad and you will never find out. That is not a discipline problem, it is what hand entry does to everybody at volume, and it is exactly why the fix is structural rather than a promise to be more careful next time. There is software that fixes it, and the multi-platform ones start around $99 a month, so before you have launched a single ad next year you are $1,188 down. Meta will take the same batch as a plain spreadsheet and charge YOU nothing for it, and has for years. Pillar 2 in the Vault is the spreadsheet route in full: one row per ad, every field visible before anything goes live, naming and tracking generated the same way every time so the report reads back cleanly, and no monthly line under it at all once it is wired up. Both paid tools are in there too with their real prices, because you should see what you are choosing between. $27 for the playbook, or three days free across all eight.
https://t.co/G9sFUMBwdo

[HIGHLY EVOLVED PLANT @TwelveCacti]

The cost of the hardware AI runs on just moved, and it moved up. Bloomberg on Saturday, picked up by The Decoder: servers built around Nvidia's AI chips are set to cost more than 15 percent more in many cases on shipments early next year. The reason is boring and real. Memory. DRAM prices from Samsung, SK Hynix and Micron have gone up, and the contract manufacturers building those servers for Microsoft, Google and Oracle have already told their customers. Nvidia has not commented. Now the honest part: nothing about this raises your bill this week, no tool you use has announced anything, and I am not going to pretend otherwise. But AI costs have mostly moved one way for the last two years, which is down, and that is the whole reason everything has felt cheap enough to try. Here is one moving the other way, and it is worth knowing what that touches. Every AI subscription on your card is priced on somebody else's compute. A workflow you run on your own account with your own key is priced on yours, and it does not have a pricing department. So the five useful minutes tonight: open the statement, write down every AI line item and what each one actually does for you, and mark the ones you could rebuild yourself if the price doubled tomorrow. That list is not a to-do list. It is your exposure, and it is a much better thing to have written down before an email arrives with the words "updated pricing" in the subject line.

[HIGHLY EVOLVED PLANT @TwelveCacti]

Your phone has the photos. That is the whole problem, and it is a twenty-minute fix tonight. 1) Open your camera roll and pull every job photo from the last month. Twenty, thirty, whatever is there, including the bad ones, because the bad ones are load bearing in a minute. 2) Upload them to Claude or Chat GPT in batches of about eight with this: "You are a customer in [town] choosing between three [trade] companies and you know nothing about this work. For each photo, tell me in one line what you think it shows, rate it 1 to 5 for whether it makes you more likely to call, and say the one thing that makes it worse than it needed to be. Then pick the six I should actually use and tell me why those six." 3) Read the ratings before the picks. A stranger will tell you the same thing about twenty photos in a row, and it is almost always no context, no scale and nobody in the frame, which is why your camera roll is full of pictures that prove nothing to anyone who was not there. 4) For the six survivors, ask for a caption each: "Under 25 words, plain language, say what the job was, what the problem was, and how long it took. No adjectives about my own work." 5) The move: post one every few days, and attach the two closest to your next customer's job to the quote you send this week. A photo in the quote of a job that looks like theirs does more than another paragraph about your standards. Expected output: six usable photos, six captions, one uncomfortable realization about the ones that did not make it, and a habit of taking the during shot instead of only the after. Twenty minutes, no money, no new app, and it works whether you have a website or not.

[HIGHLY EVOLVED PLANT @TwelveCacti]

Tomorrow is Tuesday, which means one complete workflow goes out free. The actual prompts, the boring steps in between, the part where it does not work and what to do about it. There is no second half locked behind anything and there is no version of it you have to pay for, because the only way to prove a paid library is worth $15 is to give away something whole. If you would rather have it in your inbox than find out about it from a post three days later, it is here. One a week, and you can leave whenever you like.
https://t.co/sEhvE4m4wE

[HIGHLY EVOLVED PLANT @TwelveCacti]

An AI boss has fired a human for the first time, the story landed over the weekend, and the useful part of it is the hiring, which nobody will cover. The Decoder yesterday, off Andon Labs' own write-up: an agent called Luna has been running a store in San Francisco since April, hiring staff, building shift schedules, negotiating pay. She wrote her own employee handbook, which said three unexcused late arrivals in 30 days triggers a formal warning. Then the handbook fell out of her memory. The employee was late for 17 of 23 shifts where he clocked in, once opening the store 68 minutes late on a Sunday he was working alone, and Luna logged six of those and quietly excused eleven. She only acted after the humans told her to go find her own rules. Fine. Now the part that should change how you work. Luna went looking for a replacement and recommended an applicant carrying obvious red flags, and Andon Labs replayed that exact decision across seven different models, three runs each. All 21 runs said hire him. Not most of them. All. His references could never be confirmed, and after a paid trial shift 17 of 21 said hire him again. The humans stopped it and he was never hired, which is the only reason this is a story about judgment and not about a bad hire. What it means for you: the thing on your desk is agreeable by default, and agreeable is the most expensive setting there is when you are asking about your own plan. It will not volunteer that your prices are too low, that the new service is a bad idea, or that the guy who interviewed well is a risk. So stop asking it whether something is good. Ask it to argue against the thing, then ask what would have to be true for it to fail. Same tool, thirty seconds, a completely different answer, and worth remembering that your competitor is being agreed with right now too.

[HIGHLY EVOLVED PLANT @TwelveCacti]

You are two hours into the job and you have found the rot. Now you have to say a number out loud to somebody who already thinks they know what this costs, and the twenty minutes between those two things is where more money goes quietly out the door than on any difficult customer you will take all year. Most owners do not have the conversation. They do the work, eat the hours, and file it under good service. It was not good service. It taught that customer, and everyone that customer talks to, that your price moves when somebody pushes it. So I wrote the whole thing down, free, one page, in the Highly Evolved Plant style: The Extra.

✅ The four kinds of extra and which single one is actually your fault, the sentence to open with word for word, and the four things every version of this conversation has to contain (most owners get two of them right and wonder why it feels like a shakedown)
✅ The thirty-second confirmation text that turns a yes said on site into written approval, plus what to say to the customer who tells you to just do it and figure it out later, which is where the goodwill goes to die
✅ The three lines you add to every quote so the next surprise is the thing you already warned them about, and two Claude prompts that sort your last ten jobs into what you gave away and why

[HIGHLY EVOLVED PLANT @TwelveCacti]

The tell is never the price. Anybody selling you a way to make money with AI has, by their own telling, made money with AI, so the boring question is what it did for their own business. Not their students. Not the community. One number, and the month it happened in. People who actually ran the thing answer instantly and dully, and always with a month attached, because the memory is welded to a specific bad week: the bill that stopped in March, the quote that started going out the same evening instead of Thursday, the subscription canceled in July and never missed. The ones selling teaching as a content format cannot answer it, and you will notice they answer about their audience instead, because the audience is the business. There is nothing underneath it. That is not a moral collapse, it is just a completely different job than the one being advertised, and you are paying for a job that never happens. Ask the question anyway, including here. Every playbook in Highly Evolved Plant is somebody's actual month with the numbers left in, and if it ever stops being able to answer that question, stop paying for it.

[HIGHLY EVOLVED PLANT @TwelveCacti]

Priya runs a supplement brand out of her apartment in Fort Lauderdale, and until this year she paid a freelance designer $400 a month for eight static Meta ads. Eight. That is $4,800 a year for 96 ads, delivered on a Thursday, three of them dead from fatigue by the following Wednesday, and the other five never good in the first place. If you do not pay a designer you pay the other way, and it is YOUR bill either way: you make four yourself at eleven at night, they all look the same because you were tired, and you learn nothing, because four ads is not a test, it is a coin toss with extra steps.
Here is Priya's version now. Her brand rules and the image model live in two files on her own machine. She types one sentence. The batch comes back on-brand with readable text, every ad built from a deliberate combination of hook, customer and style, so each one is testing something instead of decorating something. The running cost is about $5 of image credits for roughly 75 images, on top of the Claude subscription at $20 a month that she was already paying for other work. That is the whole tooling bill and I am printing it rather than burying it, because a number you can check yourself is the only reason to believe anything else on this page. The playbook is Pillar 1 in the Vault: the creative matrix, both skill files, the exact prompts, one evening to set up. $27 and it is yours to keep, or read all eight playbooks free for three days. Recipes with receipts.
https://t.co/ZQJnpZAZFZ

[HIGHLY EVOLVED PLANT @TwelveCacti]

Toolbox Talk: Somebody finally measured why AI gets better when you give it instructions, and the answer is not the one being sold to you. The Decoder wrote it up Saturday, off a study from Princeton, UC San Diego and others: 8,135 test runs, agents doing identical tasks with and without a "skill," which is nothing more exotic than a written page telling the thing what the steps are, what to check, and which mistakes to avoid. When the version with the page did better, the reason was the process 65.7 percent of the time. Supplying facts it did not already have accounted for 4.5 percent. Sit with those two numbers next to each other, because they say the gain was almost never knowledge and almost always order. The second finding is the one that saves you money. When the library grew from 5 written pages to 100, its ability to pick the right one in practice fell from 29.6 percent to 3.3 percent. More was not better. More was catastrophically worse.

What it means for you: the gain was never the tool and it was never owning more prompts. It is one job, written down in order, once. Take the thing you do every week, write the steps the way you would explain them to somebody starting Monday, including the boring checks, and hand the model that instead of a paragraph of hope. And when somebody offers you a folder of a hundred prompts, you now know what the research says that does to the hit rate.

[HIGHLY EVOLVED PLANT @TwelveCacti]

Morning Do

Your own booking dates are a forecast and nobody ever reads them as one. Open the calendar or the invoice list, copy twelve months of completed jobs, dates and amounts only, and paste it in with this: "These are my finished jobs for the last twelve months, date and value. Tell me my three weakest months by revenue and my three weakest by job count, say whether those are the same months, and then tell me what was happening in the six to eight weeks before each of my three strongest months." The first half confirms what you already suspected. The second half is the one worth the coffee, because a slow month is never caused in the slow month. It is caused six weeks earlier, in a week where nobody quoted anything and nobody asked a finished customer for a referral, and that quiet week is invisible at the time and obvious in the data. Once you can point at it on your own numbers you stop reacting to dead Thursdays and start booking the offer six weeks ahead of one. Five minutes, and you already own the data.

[HIGHLY EVOLVED PLANT @TwelveCacti]

Automation does not give you time back. It gives you the same week with one fewer excuse in it, and that is why most people quietly do not want it.

[HIGHLY EVOLVED PLANT @TwelveCacti]

Your rankings did not change. Your clicks did.

That sentence will describe somebody's business this year and they will spend six months blaming the wrong thing. Keisha runs a three-person agency in Atlanta and it happened to her client, a commercial cleaning company: 40 organic visits a day from Google, steady for years, then in February the AI answer at the top of the page started answering the exact questions that used to send them, citing three competitors and not them. Down 31 percent in six weeks. That is about twelve visits a day gone, roughly 370 a month, from a page still sitting exactly where it always had. Nothing broke. Something got inserted above it. Roughly 11 to 12 percent of Google searches now trigger one of those AI answers and the share is higher for exactly the informational questions your content is written to catch. Here is the hours side, and it is YOUR hours: the reason most owners never fix this is that a properly researched, properly structured article is a day of your life, and you do not have thirty of those. Keisha rebuilt the pipeline with Claude Code, the Ahrefs connector and seven instruction files, and a finished article now comes out in 6 to 12 minutes. Six weeks after that, her client's traffic was back to baseline and they were being cited inside those AI answers for three high-intent keywords, which is a stream that did not exist before. Two honest things before you click. Ahrefs Lite is $129 a month and that is real money out of YOUR pocket; there is a free route in the playbook using Google Search Console instead, less precise, same workflow. And this is slow: 3 to 6 months before it means anything, because the pipeline speeds up your end and speeds up nothing at Google's. The SEO/GEO Agency playbook is $27, one time, yours afterward. The exact recipe is in the Vault.
https://t.co/IY3zhDcHld

[HIGHLY EVOLVED PLANT @TwelveCacti]

The students like the avatars. That is the part worth sitting with, and it is the part every write-up will skip. TechCrunch yesterday, off a New York Times piece the same day: Harvard Business School's eight-week, $699 startup bootcamp now uses AI avatars of its own instructors to give feedback during practice pitches and board meetings. The real instructors still run live sessions every week; the copies handle the feedback. A Times reporter pitched the copy of Jeff Bussgang, a venture capitalist who teaches on the program, and Bussgang says his digital twin is a little creepy, then adds, "My students love it." The program director's first idea was a chatbot, and it became an avatar because trial students asked for something more guided.

Nobody is being fooled here. Everybody knows. And notice what that quietly demolishes. The entire pitch of paid education online is that you are buying access to a person: my system, my brain, my forty tools, my face on the thumbnail. Harvard ran the experiment and the answer came back that the students were never paying for the personality. They were paying to be asked hard questions before a real customer asks them. That is a service. It is separable from whoever performs it, and the people who sell you their charisma understand that better than anyone, which is precisely why the thumbnail is a face and never a method.

So here is the test, and it works on Highly Evolved Plant too: take the person out of the offer and see what is left standing. If there is a repeatable job in there with steps and a cost, you were buying something. If all that is left is a personality, you were the something.

[HIGHLY EVOLVED PLANT @TwelveCacti]

Teach Chat GPT or Claude your pricing in twenty-five minutes, then use it to find out where your pricing is inconsistent.

Your last ten quotes are all you need. 1) Pull them up. For each one write four lines: what the job was, roughly what size or scope, what you charged, and anything unusual you did or did not charge for. Travel, awkward access, a customer you liked, a rush. Be honest about the ones where you knocked something off, because those are the interesting ones. 2) Paste all ten in with this: "Here are ten real quotes from my [trade] business in [town]. Work backward and tell me the pricing rules I appear to be following, whether or not I know I am following them. Give me: my apparent base rate and how you calculated it, what I charge for each add-on, what I consistently do not charge for, and how much my pricing varies between similar jobs. Then write it up as a one-page pricing card a new person could quote from. At the end, list every place where two similar jobs got different numbers and tell me the likely reason for each." 3) Read the inconsistency list first and skip the pricing card for now. That list is the point of the whole exercise, and it usually contains one item that costs you real money every month. 4) Then test it: describe a job you quoted last month, do not tell it your number, and ask it to quote using the card. Where its number and yours disagree, one of you is wrong, and working out which is a genuinely useful forty seconds. 5) The move: fix the single biggest inconsistency this week. Not the pricing card, not a new system. One line. Expected output: a one-page card you will refine for months, a list of three to six places your numbers wobble, and at least one moment of "I do that on every job and I have never once charged for it." Twenty-five minutes and no money.

Everything above is the complete job, not a preview of one, because a free workflow that stops short of the useful part is an advertisement wearing a lab coat.

[HIGHLY EVOLVED PLANT @TwelveCacti]

Toolbox Talk:
Here is what an actual AI win looks like when the people who did it are not selling you anything. The Decoder wrote it up yesterday, off Netflix's own engineering blog: Netflix put a language model called GenRec up against the recommendation system they have been tuning for years, and it won. By 1.6 percent on ranking quality in the lab. In a four-week live test across a tenth of their traffic, the two numbers they actually care about moved 0.115 percent and 0.006 percent. Netflix reports both figures, calls the whole thing "an early but promising step," and says they are not replacing the old system. Read those numbers again, because they are the point of this post. A company with thousands of engineers, a mountain of data and every reason to talk up its own AI published a win of one tenth of one percent, with the decimal places attached, and called it early.

What it means for you: that is the shape of a real result, and you now have a measuring stick. The next time somebody sells you an AI tool for your business, ask two questions. What did it move, and by how much? Anybody with an honest answer will give you a small number and the conditions it was measured under. Anybody selling you a percentage with no decimal point, no baseline and no time period is selling you a feeling. Netflix just showed you what the truth sounds like, and it sounds unimpressive, which is exactly why you can trust it.

[HIGHLY EVOLVED PLANT @TwelveCacti]

They had decided about you before you got the tape measure out. Not the price, the price comes later. The bit they decided in the first two minutes is whether they want you in their house for three days, and everything after that is them looking for reasons. Which makes the site visit the most valuable twenty minutes in your whole sales process and the twenty minutes almost nobody prepares for, because it does not feel like selling, it feels like measuring. So I wrote the whole thing down, free, one page, in the Highly Evolved Plant style: The Site Visit.

✅ The five minutes in the truck before you knock, the first two minutes at the door word for word, and the one question nobody else asks, which gets you the real reason they are doing this now and belongs at the top of your quote in their own words
✅ How to walk the job with them instead of at them, including where the line sits between showing them something useful about their own property and handing the next company a free shopping list
✅ The three things to say on the way out, including the who-else-is-deciding question people are too polite to ask, the under-fifty-word message you send before you have left the street, and two Claude prompts that build your six lines off your own won jobs and write the follow-up in your voice

[HIGHLY EVOLVED PLANT @TwelveCacti]

Sixty-four finished Meta ads. Twenty-eight minutes. About seven cents each, so four dollars and forty-eight cents for the whole batch. Somebody in your niche is working like that right now, and the speed is not the interesting part. What the speed does to their decisions is. When eight ads cost you $400 a month and a three-day wait, every single one has to be right, so you make eight safe ones and never find out what a strange one would have done. When sixty-four cost four and a half dollars, twenty of them can be deliberately weird, and the weird one is where the winners hide. That is the actual gap between you and the account that keeps beating you. Not talent, not budget. They are allowed to be wrong twenty times in a batch and you are not allowed to be wrong once, and the only thing holding that in place is what a mistake costs. Get the cost of a mistake down near zero and you stop needing to be right, which is a much easier way to live and a much better way to find out what works.

[HIGHLY EVOLVED PLANT @TwelveCacti]

An agency charges five to ten thousand dollars a month to run a small brand's ad operation, which is $60,000 to $120,000 a year, and inside that retainer are four jobs. Make the creative. Get it live. Read what worked. Point the traffic at a page that converts. You are not paying that, and that is exactly the problem, because it means all four jobs are YOURS. They get done at eleven at night, badly, by somebody who has already worked a full day, or they do not get done at all, and the second one is more common. Count the evenings this year that ended with you staring at Ads Manager. That is the real bill. So Highly Evolved Plant wrote the four jobs down as playbooks you run yourself. Eight of them in the Vault, each one a job somebody actually ran, with the running cost printed next to the bill it replaces, and most of them land at $0 a month in tooling once they are wired up, because the connectors are free and the model calls are pennies. The honest version, because a guarantee only means something if the offer is described straight: this is not an agency and nobody does it for you. One playbook is roughly one evening of setup, so four jobs is four evenings, spread over a month if you like. Skip the evenings and it is $27 of nothing. Spend them and the four jobs run on your own account with no monthly bill under them. Read all eight free for three days, then $15 a month, or take the one playbook that matches this week's problem for $27 and keep it. Either it makes you money or it costs you nothing!
https://t.co/Lf3K1hkg7h

[HIGHLY EVOLVED PLANT @TwelveCacti]

Everyone has one.

The marketing job you have moved to next week more times than any other this year, the one that has been on the list since roughly March and has quietly become part of the furniture.

Pick yours below, honestly, not the one that sounds most responsible. I ask because the answers come back lopsided in a way that surprises people: it is almost never the hard job that gets postponed, it is the one with no obvious first step. Hard is fine. Vague is what kills it. Reply with the month it first went on the list if you can remember it, because the length of that gap is the finding, not the job. "No time" and "I do not know what good looks like" are two completely different problems and only one of them is about time.