ChatGPT, A to Z

Creating your account

Marco has run his plumbing shop in Columbus for fourteen years. His parts supplier mentioned ChatGPT at the counter one Tuesday morning, and Marco figured he'd take a look that evening, mostly to prove to himself it wasn't worth his time.

On his phone he went to chatgpt.com and tapped Sign up. It offered to continue with Google, which meant no new password to forget, he used the same Google account he already uses for his business email. A minute later, after a quick email verification, he was staring at an empty chat box with a blinking cursor.

The free plan asked for nothing except that verification. Marco learned there is also a paid plan, Plus, at twenty dollars a month, which answers faster during busy times and unlocks the newest features first. He decided to start free. His rule for any new tool: it has to earn its keep before it gets his money. Three weeks later, after using it every single day for quotes, emails, and troubleshooting, he upgraded without blinking.

He also installed the ChatGPT app from his phone's app store and signed in with the same account. Everything synced, a chat he started on the shop laptop showed up on his phone in the truck. That mattered more than he expected, because most of his thinking happens between jobs, not at a desk. The app even has a voice mode, so he can talk through a problem out loud while driving and get a spoken answer back.

One thing worth knowing early: the account is the same whether you use the website or the app. Sign in once, and your conversation history follows you everywhere. If you ever switch phones or computers, you just sign in again, nothing is lost.

The whole setup took Marco about five minutes, most of it waiting for the verification email. By eight that evening he was asking it his first real question, and by eight-fifteen he was annoyed he hadn't tried it a year earlier. Don't overthink this step, the learning starts the moment you send your first message, and you can't break anything by asking.

The interface tour

Priya owns a bakery and does everything on a laptop wedged between the register and the ovens. When she first opened ChatGPT, she spent ten minutes just looking at the screen, afraid to click the wrong thing. Here's the tour she wishes someone had given her.

At the bottom of the screen is the big message box, that's where you type. To its left is a plus or paperclip button for attaching things: a photo of a cake a customer wants replicated, a PDF of your price list, a spreadsheet of expenses. To its right is a microphone icon, tap it and dictate instead of typing, which Priya now does while her hands are covered in flour.

At the top of the screen is the model picker. Think of it like choosing which employee to hand a task to, the default option handles everyday work well, and you rarely need to touch this as a beginner. Don't let it distract you in week one.

The left sidebar is where the real organization lives. Every conversation you start is saved there as its own thread, with an automatic title. Priya's sidebar now reads like a filing cabinet: "Saturday market captions," "Supplier email, flour prices," "New menu descriptions." She starts a fresh chat for each new topic, because each thread remembers only its own conversation, which is exactly what you want. A chat about birthday cakes shouldn't bleed into a chat about hiring.

Two small tricks she uses daily: hover over any chat to rename it to something she'll recognize, and the "New chat" button at the top of the sidebar is her most-clicked button in the whole app. When a thread gets long and the AI starts forgetting the beginning, she just starts a new one and pastes in a short summary.

Her takeaway after week one: the interface is simpler than it looks. One box to type in, one sidebar that files everything for you. The skill isn't learning buttons, it's learning to keep your topics in separate threads so every conversation stays sharp.

Your first 10 prompts

Denise owns a hair salon and is methodical about everything, so instead of poking at ChatGPT randomly, she wrote herself a list of ten real jobs from her week and ran each one as a prompt. She calls it her reconnaissance mission: the goal wasn't perfect results, it was learning where the tool is strong for her business.

Her ten: a Facebook post for a slow Tuesday promotion, a reply to a three-star review, rewording her color-service price list in friendlier language, five interview questions for hiring a junior stylist, a text reminder for clients who no-show, answers to common booking questions, an email to her product supplier about a late delivery, a birthday-month promo idea, a description for a new scalp treatment, and a plain-language summary of a long article about salon retail trends.

Seven of the ten came back genuinely useful. The price list rewording was so good she put it on her booking page the same day. Two were mediocre, the birthday promo was bland, and one was useless, because she gave it almost nothing to work with. That spread taught her more than any tutorial could: ChatGPT shines at drafting words, organizing thoughts, and brainstorming options, and it struggles whenever the prompt is vague.

She saved the seven winners into a note on her phone titled "prompts that work." Two months later that note has forty entries, and it's the most valuable document in her business that she didn't pay anyone to write.

Your turn: don't aim for brilliance on day one. Aim for ten attempts across ten different jobs. Notice which ones save you real time, those are the workflows you'll build on in Module 3. Notice which ones flop, and ask yourself what information was missing from the prompt. That instinct, spotting the missing ingredient, is the entire skill this course teaches.

Keep every thread, good and bad. Next month you'll look back at your first prompts and laugh at how far your prompting has come, and that progress will be visible right there in your sidebar.

Being specific

Ray runs an HVAC company, and his first prompt to ChatGPT was "write an ad for my business." What came back could have advertised anything, a pizza shop, a law firm, a yoga studio. He almost closed the tab. Then his son, home from college, leaned over and said: "Dad, you basically told it nothing. Watch this."

His son typed a prompt with four parts. A role: "You are a direct-response copywriter for home service companies." A task: "Write a Facebook ad." Context: "We do residential AC repair in Columbus, Ohio, twenty years in business, same-day service, $79 tune-up special this spring." And a format: "Under 80 words. Headline, body, call to action."

The result read like it was written by someone who had ridden in Ray's truck. Specific words, local references, a real offer. Ray's reaction: "Okay. That's different."

The difference wasn't magic, it was information. ChatGPT can only work with what you give it, and business owners already know their business cold. The entire skill is learning to say what you know inside the prompt. Vague in, vague out. Specific in, specific out.

Ray's shortcut now: before any important prompt, he asks himself four quick questions, who should it act as, what exactly should it make, what does it need to know about my business, and what should the finished piece look like? Ten seconds of thinking, dramatically better output. He keeps the four-part structure taped to his monitor, and his office manager has started using it too. The format part is the one most people skip, and it's the one that saves the most back-and-forth: say "under 80 words" and you get 80 words, not 300 you have to trim yourself.

Try both versions yourself on any task this week, the three-word version and the four-part version, and compare. You'll never go back to the three-word version again.

Giving context

Sofia manages the front office of a dental practice, and she was tired of re-explaining the practice at the start of every chat. "We see mostly families, lots of nervous patients, warm reassuring tone, never quote exact prices...", typing it every time felt like introducing herself to the same coworker each morning.

So she wrote a context paragraph once, saved it in a note on her phone, and pasted it at the top of important tasks. Five or six sentences: what the practice does, who the patients are, the service area, the tone (warm, calm, never clinical-cold), and two things it must never claim, specific prices and same-day availability. Suddenly the appointment reminders sounded like her office, the social posts fit nervous first-time patients, and the answers stopped suggesting things the practice doesn't offer.

Then she found the permanent version: Custom Instructions. In ChatGPT's settings, under Personalization, there are two boxes, one for facts about you, one for how you'd like it to respond. She pasted her context paragraph there once. Now every new chat starts with her practice's background already loaded, no pasting required. It was the single biggest upgrade to her results, and it took four minutes.

Her rule for the context paragraph: include what the business does, who the customers are, the tone, and anything it must never say. Leave out anything you'd be uncomfortable sharing, no patient names, no financial details. Context is about the business's public identity, not its private data.

She also learned to update it. When the practice added Saturday hours, she edited the paragraph. When they hired a pediatric dentist, she added a line. Five minutes of maintenance a quarter keeps every answer current.

The pattern to remember: ChatGPT knows a lot about the world and nothing about your business until you tell it. Tell it once, in writing, and every future answer gets smarter.

Iterating: the follow-up is the skill

Tom owns a landscaping company, and every March he dreads writing the spring cleanup flyer. This year he asked ChatGPT for a first draft, and what came back was too long, too stiff, and missing his prices. The old Tom would have sighed and written it himself. The new Tom typed: "Make it half as long and friendlier."

What came back was closer. "Add our $199 starting price and a line about gutter cleaning." Closer still. "Give me three headline options for the top." He picked the second one. Four follow-ups, six minutes total, and the flyer was done, better than the one he sweated over for two hours last year.

The key: he stayed in the same chat thread the whole time. Each follow-up built on everything before it, because the thread remembers the conversation. If he had started a new chat for each revision, he'd have been re-explaining the flyer from scratch every time. One thread per project, start to finish, that's the habit.

He also learned two thread tricks. First, you can edit one of your own earlier messages, hover over it, click edit, change the wording, and ChatGPT will regenerate everything after it. Useful when you realize your original instructions were the problem. Second, if an answer is close but not quite, you can hit regenerate for a fresh take on the same prompt instead of rewording it yourself.

Tom's mental model now: the first answer is raw material, never the finished product. He treats ChatGPT like a junior employee, brief it well, then revise. "Shorter." "More professional." "Add a headline." "Now make a version for returning customers." Each follow-up is cheap, fast, and compounding.

The business owners who get the most from ChatGPT aren't better writers, they're better revisers. They expect draft one to be imperfect and they know exactly which follow-up sentences fix it. After a week of practicing, you will too.

Writing social posts

Rosa's bakery does a farmers market booth every Saturday, and every Friday night she used to stare at her phone for twenty minutes trying to write a post about it. The croissants were ready at 5 a.m.; the caption was never ready at all.

Her new Friday routine takes six minutes. She opens the same chat thread she uses every week, it already knows her bakery, her tone, and what sells, and types: "Write 5 Facebook captions for tomorrow's market booth. Warm, appetizing tone, under 40 words each, end each with a question to drive comments." She picks the two she likes, changes a word or two so they sound like her, and schedules them.

Three things made the difference. First, she asks for several options at once, five captions, not one, because choosing among options is faster and better than revising a single draft. Second, she specifies the platform, because a Facebook caption and an Instagram caption are different animals. Third, she never posts the raw output. The AI drafts; Rosa decides. That last rule is non-negotiable in her shop.

She also batches. Sunday night, one thread, seven days of posts: "Give me one post idea per day next week for the bakery, Monday behind-the-scenes, Wednesday new item, Friday market reminder, and fill the rest." Twenty minutes, and her social media is done for the week. Before ChatGPT, that same job ate two hours she didn't have.

When she needs a promo graphic, she asks ChatGPT to create the image too, it can generate pictures right in the chat. A "Saturday market" graphic with croissants and her booth colors takes one prompt. It's not a professional designer, but for a quick social post it's more than enough, and it beats a blurry phone photo.

Her takeaway: social media stopped being a chore when she stopped writing from a blank page. The blank page is the enemy; five options to choose from is a gift. Batch it weekly, pick fast, tweak lightly, post consistently.

Email drafts

Marco the plumber is good with his hands and terrible with follow-up emails. He knew he was losing jobs, a customer gets a quote, goes quiet, and Marco means to follow up but never finds the words. The quote sits. The customer calls someone else.

His breakthrough was embarrassingly simple: he pastes the customer's original email into the chat and writes, "Draft a polite follow-up to this. Friendly, zero pressure, invite questions, under 120 words." What comes back needs two small tweaks and it's ready to send. The whole thing takes three minutes, which means it actually gets done.

Pasting the original message is the secret to this whole lesson. When ChatGPT can see what the customer actually wrote, their wording, their concerns, their tone, the draft comes back sounding like a real reply instead of a form letter. Marco now does this for quote follow-ups, rescheduling notices, "we're running 20 minutes late" messages, and replies to supplier emails about pricing.

He has one iron rule: he reads every draft before sending. It's his name at the bottom, not the AI's. Twice now the draft has included a detail that was slightly off, a wrong day, an assumed price, and he caught it because he read. The AI drafts; Marco signs. That division of labor is the whole point: it does the hard part (starting from blank), he does the easy part (checking).

He also keeps a thread called "email drafts" so he can scroll back and reuse structures that worked. His "sorry we're late" template has now been used eleven times, each slightly reworded for the situation. The thread is basically a personal swipe file that writes itself.

If email is where your business leaks money, unanswered quotes, slow replies, awkward wording, this is the highest-ROI lesson in the course. Three minutes per email, jobs you were already losing now get followed up.

Ad ideas and headlines

Ray's HVAC company spends real money on Google Ads, and the headline limit, thirty characters, spaces included, used to defeat him. "AC Repair Columbus" is 18 characters and says nothing. He'd write three options, pick the least bad one, and wonder why the phone wasn't ringing.

Now he asks ChatGPT for ten headlines at a time, with the constraint built into the prompt: "30 characters max each." It counts characters better than he does, and ten options take thirty seconds. He picks three, runs them against each other in Google Ads, and keeps the winner. Last month the winning headline, "AC Out? Fixed Today", beat his old one by enough that he noticed it in the invoice.

The deeper lesson wasn't the headlines; it was the angles. Ray asked for "three completely different approaches: one about price, one about speed, one about trust." The price angle flopped. The trust angle ("20 Yrs Fixing AC Right") won. He would never have tested the trust angle on his own, he'd have kept shouting about discounts. The AI is a tireless brainstorming partner; Ray remains the judge.

He uses the same approach for Facebook ads, where the constraint is different, there he needs a hook in the first line, because everything below "see more" is invisible. "Give me 5 opening lines that stop a homeowner scrolling past an AC ad in July." Summer in Ohio writes the urgency for him; the AI writes the words.

His rule: the AI generates options, never decisions. It doesn't know his margins, his close rate, or which jobs he actually wants. Ten ideas in, three tested, one winner, and the testing, not the AI, is what makes the money.

Start with whatever you're already running. Paste your current ad, ask for ten alternatives under the platform's character limit, and test the best three. You'll learn more from one week of testing than from a year of guessing.

Drafting replies

Denise found the two-star review at 11:40 p.m., twenty minutes after a fourteen-hour day. "Waited 45 minutes past my appointment, won't be back." Her thumb hovered over the reply button, and every word she wanted to type would have made it worse.

Instead she screenshotted the review, opened ChatGPT, and pasted it in: "Write a calm, professional reply. Apologize for the wait, thank them for the feedback, invite them back with 20% off their next visit. Keep it under 80 words." What came back was gracious in a way she couldn't have managed at midnight. She saved it, slept, re-read it in the morning, changed two words, and posted it. The reviewer never came back, but three new clients mentioned the professional reply when they booked.

That's the real use case for this lesson: draft when calm, send when ready. The AI doesn't get defensive at midnight. It doesn't take reviews personally. It just produces a professional starting point, and you supply the judgment.

Denise now uses the same move for direct messages. "A client is running 15 minutes late and my next appointment is waiting, what do I text each of them?" She gets two drafts, personalizes them, sends. For the rare truly angry message, she asks for "three versions: warm, firm, and brief," then picks the tone the situation needs.

Her rules, learned the hard way: always personalize before posting, a reply that sounds copy-pasted is worse than a slow one. Never post a reply you haven't re-read. And never, ever let it invent facts: if the draft says "we've updated our scheduling system" and you haven't, delete that line. The AI drafts words; only you know what's true.

Reviews are permanent and public. A good reply is marketing that money can't buy, and now it takes Denise five minutes instead of a sleepless night.

FAQ documents

Sofia's dental front desk answers the same questions all day: Do you take my insurance? How much is a cleaning? Do you see kids? Are you open Saturdays? Her team answered them patiently, one phone call at a time, hundreds of times a month.

She taped a notepad to the front desk and asked the team to write down every question they heard for one week. Sixty-three questions. A lot of repeats, which was exactly the point, the repeats are the FAQ.

She pasted the list into ChatGPT: "Turn these into a clean FAQ page. Group by topic, insurance, pricing, appointments, services. One short answer each, plain language, warm tone." Ten minutes later she had a document. She spent another twenty fixing the details only she knows, the exact insurance plans, the real Saturday hours, because the AI can organize answers but can't know your facts. That's the division of labor for this whole module: it structures, you verify.

The FAQ went three places: the practice website, a printed sheet in the waiting room, and a pinned thread in the team's group chat for new hires. Phone volume on routine questions dropped noticeably within a month, Sofia didn't measure it precisely, but the front desk noticed, and the front desk notices everything.

Her maintenance habit: every quarter she asks the team "any new questions lately?" and pastes the new ones into the same thread, which still remembers the original document's style, and asks for an updated version. The FAQ grows with the business instead of going stale. New hires get the pinned thread on day one, which cut Sofia's training time for front-desk staff nearly in half: instead of shadowing for two weeks to absorb the answers, they read the FAQ in an afternoon and shadow for confidence.

Your business already has an FAQ; it's just stored in your employees' heads and repeated verbally a thousand times. One week of collecting, one afternoon of organizing, and it works for you around the clock.

Tone matching

Tom's landscaping texts sound like Tom: "hey! we'll be there tmrw morning 👍 see you then." Short, warm, a little rough around the edges, and his customers love him for it. The problem was that ChatGPT's drafts sounded like a bank: "Dear valued customer, we wish to inform you of our scheduled arrival..." His customers would have thought he'd been replaced by a robot.

The fix took one message. Tom pasted three of his own texts into the chat and wrote: "Write all my customer messages in this same casual, friendly style from now on." The next draft came back sounding like him, "hey! quick heads up, we're running about 20 min behind, see you soon 👍". He laughed out loud in his truck.

This is the most important technique in the customer-service module: show, don't describe. Telling the AI "be casual" produces generic casual. Showing it three real examples of your casual produces your voice. Adjectives are vague; examples are precise. Two or three samples is enough, it picks up rhythm, word choice, even emoji habits frighteningly fast.

Tom now starts every customer-facing thread by pasting his style samples first. He keeps them in a phone note titled "my voice," right next to his context paragraph from Module 2. New thread, paste both, and every draft comes out sounding like Tom on his best day. The real test came when a customer replied "haha sounds good man" to one of the AI-drafted texts, Tom knew the voice match was working when nobody could tell the difference, including him on a busy morning.

Test it yourself: generate a draft the normal way, then generate one after pasting your samples, and read them side by side. The difference is the difference between a template and a person. Your customers can feel it too, and in a service business, feeling like a person is the whole brand.

What a GPT is

Rosa was scrolling the ChatGPT sidebar one evening when she noticed a button she'd been ignoring for weeks: "Explore GPTs." Curious, she clicked it and found a store full of pre-made assistants, a logo designer, a recipe converter, a social media helper. She tried the logo designer, typed "bakery logo, wheat, warm colors," and got four options in a minute. Fun, but the real insight came a minute later.

A custom GPT is ChatGPT with three upgrades: saved instructions that apply every time, optional reference files it can consult (a price list, a menu, a policy doc), and a shareable link so other people can use it. Think of a regular chat as a conversation and a GPT as a tool, one you configure once and then hand to anyone.

The difference that matters for a business owner: a chat thread forgets its job. Every new chat starts blank, and you re-explain everything. A GPT remembers its job permanently. Rosa's "Market Booth Helper" GPT knows her products, her prices, and her tone, she opens it Saturday morning and just types "today's specials post," no setup, no context paragraph, no re-explaining.

GPTs also come in two flavors: ones other people made (the store), and ones you make yourself (next lesson). The store is worth browsing for ideas, Rosa found a "Canva design helper" that improved her flyers, but the real power is building your own, because only you know exactly how your business works.

One practical note: creating custom GPTs requires the Plus plan. Rosa did the math, twenty dollars a month for a tool her whole team uses daily, and it paid for itself the first week. But finish this lesson first and build your first one before deciding; you'll know within an hour whether it's worth it for you.

Your takeaway: chats are conversations you start over each time; GPTs are assistants you train once. When you catch yourself pasting the same instructions for the fifth time, that's a GPT waiting to be built.

Building one for your business

Marco's dispatcher was drowning in Facebook messages: "how much to fix a leaking faucet?", "do you come to Dublin?", "my water heater is making a noise, what do I do?", each one needing a human to triage. Marco decided to build a custom GPT to do the first pass. Total build time: twenty-five minutes. Here's exactly what he did.

He clicked "Explore GPTs" in the sidebar, then "Create." The builder opens in a conversational mode, it literally interviews you: "What should your GPT do? Who is it for? What tone should it use?" Marco answered in plain words: "It helps plumbing customers describe their problem before we arrive. Friendly, practical tone. Always ask what's leaking, where, and how urgent it is. Never state final prices, say a technician will confirm on site."

Then came the two power features. First, instructions: in the Configure tab he wrote the rules the GPT must always follow, ask for a description and photos, suggest shutting off the water valve for active leaks, never diagnose definitively, always end by offering to book. Second, knowledge files: he uploaded his one-page price sheet and his service-area list using the upload button, so the GPT answers "do you come to Dublin?" from his actual document instead of guessing.

Then he tested it like a skeptical customer. He opened the preview panel and typed the three hardest messages his dispatcher receives. Two were perfect. One gave a price estimate, against his rules, so he went back to the instructions, added "NEVER give price estimates, even ranges," and tested again. Fixed. That test-fix loop is the whole craft of building GPTs: try it, break it, tighten the instructions.

He named it "Quote Helper," saved it, and got a share link. His dispatcher now pastes tricky customer messages into it for a first-draft triage, and Marco is training his after-hours answering service to use it too. Twenty-five minutes of setup, and it handles the repetitive first pass on every inquiry.

Your first GPT doesn't need to be ambitious. Pick the one repetitive question-answering job in your business, build it the way Marco did, interview, instructions, one knowledge file, test, tighten, and you'll understand the pattern forever.

Sharing with your team

Ray built a GPT that drafts professional update texts for customers while his techs are on site, "your new unit is installed, here's what we did, call us if anything seems off." It worked great for Ray. Then he realized the real win: his two technicians could use it too, straight from their phones, and every customer would get the same quality update.

Sharing is one click: open your GPT, hit Share, copy the link. Anyone with the link can use it, no setup, no instructions to forward, because the instructions live inside the GPT itself. Ray texted the link to both techs with a short message explaining what it was for. They were using it the same afternoon.

But Ray also set ground rules on day one, and this is the part most owners skip. His three rules: the draft is a starting point, read it before sending, because your name is on the text. Never paste a customer's address, gate code, or payment details into it, describe the job, not the person. And it's a helper, not a decision-maker, if a customer is upset or the situation is unusual, call Ray, don't ask the GPT.

Two weeks in, Ray checked what the team was actually asking it, the questions revealed gaps in his instructions, so he tightened them. A tech kept asking it to explain warranty terms, so Ray uploaded the warranty sheet as a knowledge file. The GPT got better because real usage showed him what was missing. Build, share, watch, improve, that's the loop.

The compounding effect is what sold him: one link, whole team, same quality, zero training sessions. His techs aren't "learning AI", they're tapping a link and getting better customer texts. That's what successful business AI looks like: invisible to everyone except the owner who set it up.

When you share yours, write the rules down first. Five minutes of expectations prevents five months of bad habits.

Hallucinations explained

Denise was considering a new hair treatment line when the salesperson claimed "studies show" it reduces breakage. Skeptical, she asked ChatGPT: "Is there research behind this claim? Cite the studies." Back came a confident answer with two study names, authors, and a journal. It looked airtight, until Denise googled the studies and found they didn't exist. Not the authors, not the journal, nothing.

She hadn't been lied to. She'd been hallucinated at, and understanding the difference matters. ChatGPT doesn't look things up when it answers from its training; it predicts the most likely next words. When it doesn't know something, it doesn't say "I don't know", it generates the most plausible-sounding answer, complete with realistic-sounding citations. It's not deceit; it's fluent guessing. The confidence of the tone has zero relationship to the truth of the content.

Hallucinations cluster in predictable places: specific facts (dates, numbers, names), citations and references, obscure topics, and anything after its knowledge cutoff. It's least trustworthy exactly where it sounds most authoritative, a smooth, detailed, specific answer about something you can't verify on the spot.

Denise's new habit is simple: the more specific and consequential a factual claim is, the more she verifies it. A brainstorming list? No verification needed. A statistic she's putting in an ad? Verified before it goes anywhere. She also learned to ask directly: "How confident are you in this answer, and which parts should I double-check?", it's surprisingly honest when asked. The model will often point straight at the shakiest sentence, which saves her from checking everything and lets her focus on the one claim that matters.

The takeaway isn't fear; it's calibration. ChatGPT is a brilliant drafter and an unreliable witness. Use it for the first job, never the second, and the next two lessons give you the exact verification habits that make it safe. Once Denise internalized that single distinction, every other lesson in this module clicked into place.

Verifying important facts

Sofia's rule at the dental practice is taped to her monitor: "If it goes public or involves money, verify it." The practice's website, ads, invoices, and patient handouts all pass through this filter, and since ChatGPT now drafts most of them, she built verification into the workflow instead of treating it as an afterthought.

Her method has three steps. First, she asks ChatGPT to flag its own uncertainty: "List the factual claims in this draft and tell me which ones you're least sure about." It's remarkably cooperative when asked directly, it'll point at the exact sentences that need checking. Second, she checks business facts against her own records: prices, hours, insurance plans, service lists. The AI can't know these; only she does, and this is where most errors hide. Third, for outside facts, a statistic, a regulation, a medical claim, she checks a real source, not the AI's word.

She also uses a prompt trick for high-stakes questions: "Only answer what you're sure about. If you're guessing, say so clearly." This doesn't eliminate hallucinations, but it dramatically reduces the confident-and-wrong answers, because it gives the model permission to be uncertain, something its default style discourages.

The categories she always verifies: numbers (prices, percentages, dates), names (people, products, places), and claims about the outside world (laws, studies, "most dentists recommend"). The categories she doesn't bother with: brainstorming lists, first drafts of tone and structure, and anything she'll rewrite heavily anyway.

Verification sounds slow, but Sofia timed it: checking a draft takes five minutes; fixing a published error takes five hours and costs trust. The cheapest quality control in her business is a five-minute habit.

Make the rule yours this week: before anything AI-assisted goes public or touches money, run the three steps. You'll catch something within the month, everyone does, and you'll be glad you looked.

What not to trust it with

Tom's most expensive ChatGPT lesson came from a contract clause. A commercial client sent over a maintenance agreement, Tom pasted a paragraph into the chat and asked what it meant, and the answer sounded completely reasonable, until his lawyer reviewed it and found the AI had missed a liability shift that would have cost Tom thousands if something went wrong. The answer wasn't gibberish; it was confidently incomplete. That's the dangerous kind.

Tom's rule now, and yours: ChatGPT is a brilliant assistant and a terrible professional. It can explain a contract clause in plain words so you walk into your lawyer's office informed, it cannot replace the lawyer. Same for tax questions, medical guidance, and employment law. Use it to get smart enough to ask good questions; make the final call with a licensed human.

The second line is privacy. Tom's techs once pasted a customer's full name, address, and gate code into a chat to draft an arrival text. Tom stopped that the same day. His policy: describe the job, never the person. No customer names, addresses, phone numbers, or payment details in any AI chat, ever. Drafts can say "your technician arrives at 9" without knowing whose house it is.

The third line is final decisions. ChatGPT is excellent input, options, trade-offs, second opinions, and a bad decider. Tom asks it to argue both sides of buying a new mower; he makes the call. He asks it to draft the price increase letter; he decides the number. The pattern: AI informs, human decides.

None of this diminishes the tool, it defines it. A chainsaw is incredibly useful once you respect what it's for. Use ChatGPT for everything in this course, keep a professional for the three lines above, and you'll get all of the upside with none of the expensive lessons. Tom's summary to his brother-in-law, who asked if AI was "safe for business": "It's the best employee I've ever had, as long as I remember it's still an employee, I check the work."