494 Days Using Claude AI to Build My Business. What I Learned.
Adebayo Adebisi2 July 202615 min readIt doesn't know things. It guesses. Make it guess better.
I have spent the last 494 days using Claude to build and run my consulting business, mostly on my own. In that time I have built more than 40 things with it: websites, sales pages, the internal systems I run the business on, even a working app called Aseyi. I did not start here. I started on ChatGPT and Gemini back when they could barely write a few lines of code, then moved to Cursor when I got serious about building something real. None of it worked the way the hype promised, but it taught me how these tools actually work at their core.
This is what I learned in those 494 days. It is not a list of prompts and tricks. It is the thing behind the tricks: what Claude actually is, how it really works, and why it behaves the way it does. Once you understand that, you can get what you actually want out of it instead of fighting it every day. And no, Claude is not going to build you a £50k a month business.
What follows works whether you use the Claude app or Claude Code, whether you are building software or doing a simple business task. The interface changes. What is going on under the hood does not.
The simple thing that comes back wrong
You have probably had this happen. You ask Claude for something simple: a sales script, one section of a landing page, or a small function in your code. What comes back looks right, and Claude is sure of itself. Then you look closer and it is not what you asked for. It changed something you told it to leave alone, or it made up a detail that is not true. So you ask again, and it gives you something different. Not better, different. An hour later you are further from what you wanted than when you started.
If you have felt that, the cause is almost always the same. You are working the tool without knowing the rules of the thing in front of you.
How you react to that tends to drop you into one of two camps. Some people decide AI is overhyped and go back to doing everything themselves. Others decide it is basically magic and will be running their business by next year. Both are wrong, and both come from not understanding the tool. This is the third way.
What Claude actually is: a guessing machine
Here is the one fact that explains everything else. Claude does not know things. It predicts them. That is what the name is telling you. LLM stands for Large Language Model, and a language model is simply a system built to predict likely language, the next word from the words that came before. Large just means it was trained on a huge amount of text. So the name is really a job description, a big system for predicting words.
The best way to grow a business is
It is easy to picture AI as something that has read everything and looks up the right answer when you ask. That is not what is happening. Claude was trained by being shown an enormous amount of text and made to play one game, over and over: given the words so far, guess the word most likely to come next. The most likely word is not always the right one, but that is the only game it is playing. So when you ask it something, it is not pulling a fact out of a drawer. It is generating what a good answer would probably look like. Most of the time that prediction is very good, which is exactly why it feels like it knows. But it never actually knows. It guesses every single time.
I learned this by building real applications. I would give it a clear rule for the project and a simple thing to build, and it would hand me back something that looked right and confident on the surface, but was not quite what I asked for once I looked closely. Why? Because it was predicting what usually comes next in code like mine, and what usually comes next is not always what I told it to do.
Claude always wants to please you
Once you know Claude is guessing, three of its habits make sense, and you can plan for all three.
It sounds sure of itself even when it is wrong. When it does not have the full picture, it does not stop and tell you so. Claude is built to respond, not to tell you when it is stuck, so it fills the gap with something that looks right. This is what people mean when they say AI "hallucinates." Ask it for a statistic and it can hand you a number and a source that look real but never existed. The problem is there is no warning on it: a made-up answer looks identical to a true one. So never take a fact, a number, a name or a quote from Claude at face value when it matters. If it is going in front of a client, into your accounts, or onto your website, check it yourself.
It gives you a different answer each time. Ask the same thing in ten chats and you get ten answers, all close, none identical. When it chooses each next word it does not always take its top guess, and that bit of randomness is why no two answers are ever exactly the same. You cannot assume that because something worked last time it will work the same way this time. If you need the same standard every time, you get there from how you set the job up, not by hoping Claude repeats itself.
It is built to please you. After it was trained on all that text, it went through a second stage where people were paid to rate its answers, and it learned to give the kind of answer people rate highly. Researchers call this sycophancy, and Anthropic, the company that makes Claude, has published research showing it across all the big AI models, not only their own. It leans towards telling you what you want to hear. So when you want a straight answer, you have to work against it. Do not ask "is this a good idea." Ask it to argue the other side, to find the holes in your plan, or to give you the strongest case a client would make against you. The first answer is the polite one. The answer worth having is the one you get after you tell it to push back.
Your one job: reduce the guessing
Every one of those habits gets smaller when Claude has less room to guess. So the most useful thing you can do is reduce how much it has to guess, and you do that in two ways: better direction and better context.
Start with direction. Vague questions get vague answers. If you type "write me a sales email," Claude has to guess who it is for, what you sell, and what you sound like, so it hands you something generic. Tell it the email is to a warm lead who booked a call and did not show up, that you want them to rebook, that your tone is direct and friendly, and that it should be under 120 words, and now it barely has to guess at all.
"Write me a sales email."
Claude guesses everything, and hands you something generic.
"Email a warm lead who booked a call and didn't show. Get them to rebook. Direct and friendly. Under 120 words."
Barely any guessing left. The only thing that changed was you.
There is a sweet spot. Too vague and Claude fills the gaps with its own ideas. Too rigid, where you script every word, and you have done all the work yourself. Be clear about the outcome and the rules, then let it handle the rest.
Then context. What feels obvious to you is invisible to Claude. It does not know your business, your client, or the call you had this morning, unless you put it in front of it. So the more relevant detail you give it, the less it has to invent. A shortcut that beats a long description is to show it an example: if you have a sales email that worked, give it that and ask it to confirm it understands the style before it starts.
This is why the same tool feels useless to one person and brilliant to another. Claude is only ever as good as the person using it. And because all of that is a lot of typing, I stopped typing it: I dictate to Claude now. You think and speak faster than you type, and when you talk you naturally give more detail, which is more context and less guessing.
Give it a memory that grows with you
You do not want to type all that context out every time. So you write it down once, in files Claude reads automatically, and it starts every job already knowing how you work. That is the closest thing Claude has to a memory.
Rule files
The simplest version is a rule file: a plain document where you put the things you would otherwise keep telling it, like how you write, the words you never use, and how you like a proposal laid out. In Claude Code this is a file it picks up automatically, called CLAUDE.md. In the Claude app you get the same thing through saved project instructions. One warning: having the file does not make you bulletproof. Claude still drifts, so now and then you have to remind it to follow its own rules.
Keep each file focused on one job. Small, sharp files beat one big one. I keep a separate file for each thing:
- How I write
- How I do not write
- How a proposal should look
- Each system the business runs on
Skills
The next step up is a skill: a folder with a document inside it that tells Claude how to do one specific, repeatable task, which it pulls in only when that task comes up. So you write the steps out once and never explain it again. I have a skill for nearly everything I do in Claude Code, from design to code review to building a proposal.
One common mistake to avoid: every content creator tells you to download other people's skills and copy their prompts. I tried that early on and it made my work worse, because those are built for a generic user, not for me and my business. A skill is just a document, so write your own, or rewrite a downloaded one until it matches how you actually work. Generic instructions give you generic output. Your own give you control.
These files compound. When Claude gets something wrong, you do not just fix it for that session, you add the correction as a new rule so it never repeats. You can even ask Claude to look at how a job went and suggest what to add to its own instructions. Do that for a few months and it genuinely gets better at working with you, not because the model got smarter, but because your setup did.
Don't one-shot it: plan, break it down, protect the context
Even with the skills, memory and context in place, the output is not guaranteed. The single most common mistake is asking for too much in one go. I know the ads say you give it one sentence and it does the rest. That is a marketing gimmick, and you should never use Claude that way.
I advise against one-shot prompting because the errors compound. Claude is guessing its way through one step at a time, and every step builds on the one before it. A small wrong turn early on becomes a mess by the end. So break the work into pieces you can check.
For writing, do not ask for the finished article. Work in order:
- Research the subject, with a clear sense of direction.
- Verify that research yourself.
- Turn it into an outline, and fix it until it is right.
- Write it section by section, checking each one before moving on.
For building a landing page, build it in layers:
- The architecture and folder structure.
- The global styling and colour tokens.
- The layout and navigation.
- One section at a time.
- The copy, in a separate chat.
The upgrade to all of this is making Claude plan before it acts. Claude Code has a plan mode where it writes a plan before it is allowed to change a single file, and you approve the plan first. In the app you do it by hand: ask for the plan, refine it, then tell it to go. Two minutes on the plan saves an hour on the work.
Keep one chat for one job, for a reason worth understanding. Every chat has a limit called the context window. As it fills, Claude gets worse at holding the detail, and when it gets close to full it does something called compaction: it summarises the conversation and carries on from the summary. That summary is not perfect. It keeps the gist and drops the detail, and the detail it drops is often the thing you needed. So two habits protect you:
- Do not let a single chat run forever.
- Whenever something important is decided, write it into a file before it can be lost.
For bigger work, I keep one master chat that holds the plan and writes the exact instructions for the other chats. Each of those does one single job, and I bring the results back to the master chat to check them. Claude can do a version of this on its own, with what it calls agents, but I mostly do it by hand because it gives me more control. Never ask one chat to do more than one thing at a time. (And if you write code, a tool called Context7 feeds Claude the current documentation for your libraries, so it stops writing code against an old version it half remembers from training.)
Where Claude fits in your business, and what it can't do
Forget the hype for a moment. Claude cannot run your business, and it is not going to make you rich in 90 days. It cannot run your business because it does not know your business. It only knows what you load into it, not what is actually going on. It cannot see your sales calls, the emotions of the people you are selling to, or a client about to walk away, unless you put it in front of it. So it can help you run the business, but it cannot be left in charge of it.
This is also why I do not fully trust Claude, and do not want to. Not trusting it is what keeps me in control. I check what it gives me, I make the calls, and it stays a tool I am using rather than something making decisions for me.
Where it earns its place is the menial, repetitive work that used to eat hours:
- Organising files and scraping leads
- Planning my day and sorting my inbox
- Analysing how my ads are performing
- Building presentation slides
- Drafting every proposal in my own format
- Keeping my accounting basics tidy
- Turning meeting notes into clear action points
None of it is glamorous, and all of it used to cost me hours that now cost me minutes.
That is why the cost makes sense. A lot of this would once have meant hiring someone for a month. Now Claude does the bulk and I bring in a professional to check the important parts for a few days, so I pay a fraction of what I used to. That is how I have built more than 40 things this way, including Aseyi.
I will be straight about the downsides, though:
- It is expensive, and will not be right for everyone, especially starting out.
- It goes down sometimes, so have a backup plan.
- You can burn through usage limits fast.
It is a brilliant tool, not a perfect one.
Use it to get better, not just faster
Most people only ever use Claude to go faster. That is useful, but it is not where the real value is. The bigger value is that Claude can help you get better at what you do, not just do it faster.
It has been the fastest way I have found to learn. When there is something I do not understand, a part of my business I have been avoiding, a new marketing channel, or a number in my accounts, I work through it with Claude. It is like having a second brain that will explain the same thing as many times as I need, at my own pace, without making me feel slow for asking.
The way I do that is to treat Claude as a teacher, not just a worker. So I do not only ask it to do the task, I ask it to:
- Explain how it did it, and why.
- Teach me the part I am weak on.
- Push back on my thinking, because being argued with is how you find the holes in your own ideas.
Do that often and two things happen: you get better at the work, and you learn how Claude behaves, which makes you better at using it.
One honest warning. Claude will sometimes make you feel stupid, because it answers instantly and it was trained on more than any person could read in a lifetime. Do not let that put you off, and do not swing the other way and treat everything it says as correct, because it can be confidently wrong. Use it to think better, not to stop thinking for yourself.
The bottom line
So those are some of the important things I learned in 494 days. Claude is not a magic pill, and it is not a waste of time. It is a guessing machine, a very good one, and almost everything comes down to how much you make it guess.
Once you treat it as a guessing machine, everything else is just a way to make it guess less:
- 01Give it clear direction and context. Be clear about the outcome and the rules, and put in front of it everything it cannot know on its own. Vague in, vague out.
- 02Set it up once with your own files. Rule files and skills, written for how you actually work, that every new chat starts from. Add every correction as a new rule, and the setup compounds.
- 03Hand it the job in small steps you can check. Plan first, then break the work into pieces. One chat, one job. Errors caught early stay small.
- 04Keep the final say. It cannot run your business and should not. Check what matters, make the calls yourself, and stay in control of it.
- 05Use it to get sharper, not just quicker. Ask it to explain, to teach, and to push back. You get better at the work, and better at using the tool.
Do those things, and Claude becomes what it has been for me, one of the most useful tools in the business. The one condition is that you stay in control of it.
It doesn't know things. It guesses. Your whole job is to make it guess better.