Give It Context

Start here: give it four pieces of context

AI does not automatically know anything about your work. Not your prices, not your customers, not how you write. Unless you give it that information, it starts with a blank sheet of paper, and then hides that fact behind a confident answer.

The fix is not better wording. It is putting the right information in front of it. That information has a name, context, and everything in this guide comes back to it.

So start here. Four pieces of context, about an hour, no technical knowledge needed, whether you are on your own or employed.

1Tell it who you are, once.

Many AI tools let you save a note about yourself that they read every time. Look under Settings for custom instructions, personal preferences or memory. If yours does not have one, keep a short note you can paste in. Put something like this in it:

ExampleI run a small plumbing business in North Yorkshire. My customers are mostly homeowners. I write plainly and I hate jargon. Keep answers short. If you are not sure about something, say so instead of guessing.

Change it to fit you. That one paragraph stops you re-explaining yourself every single morning.

2Paste the thing. Do not describe the thing.

Want a reply to a customer? Paste their email, leaving out any private details you do not need to share. Want a quote? Paste your last good one and say “like this”. Summarising something in your own words before you hand it over is where most of the quality quietly disappears.

3Show it two things you wrote.

Find two pieces of your own writing you were happy with. Paste them in and say “this is how I write, match it.” It will copy your voice far better than any instruction about tone ever will.

4Check the numbers yourself.

It is good with words and can still get numbers wrong. Check anything that adds up with a calculator or spreadsheet, and read anything that goes to a customer before it leaves.

Before you pasteCheck what you are allowed to share, and leave out private details you do not need to.

That is it. Those four things get most people most of the value. The rest of this explains why they work, what comes next, and where it will let you down.

What it looks like when the context is missing

You ask it to write back to an awkward customer. Out comes 400 words of polite nothing. You delete most of it, write the reply yourself, and quietly wonder what everyone is so excited about.

Here is what happened. It did not know who that customer was. It did not know what you quoted them, that they have complained twice already, or that you settled it last time with a partial refund. It knew none of it. So it wrote the only thing it could: something that sounds like a reply to a customer.

Now hand it that customer’s file, the last few emails, your terms, and a dozen letters you have written before. Same AI. Same request. What comes back needs a tweak and a signature.

Nothing changed except the context.

The simple truth

Prompting helps. Context wins.

Your instructions tell it what you want. Context tells it what it needs to know. Get the second part right and the first stops mattering nearly as much.

This holds whether you are one person or two hundred. What changes is the effort of assembling the context, and for one person it is very little.

Why prompt training does not stick

Plenty of firms have sent people on prompt engineering courses. Use goes up for a month, then quietly fades. Two reasons.

No amount of clever phrasing supplies a fact it does not have. If it does not know your prices, no wording gets it there. It will guess, confidently, and the guess reads exactly like the truth.

And you never stop paying. Every job means typing the background in again. One person gets good at it and gets good results. Nobody else does. Nothing builds up. Close the window and it is gone.

Better wording is worth having. It just runs out quickly.

Six stages of context

1. Asking questions. “Write me an email.” You get an email. It could be for anyone, about anything.

2. Better instructions. You stop typing “make this better” and start typing “reply to Jim, who is furious about a late delivery, two paragraphs, apologise once, offer a firm date.” Same tool, much better answer.

3. It remembers you. You stop introducing yourself every morning. Available in many tools, often at no extra cost, and few people turn it on. This is the first of the four steps at the top.

4. It reads your own material. You hand it your documents and it answers out of them, telling you which one it used. This is the big one. Nothing needs building. In many tools you simply upload files, and some allow this on free plans. This is the step that turns “quite clever” into “that saved me a morning.”

5. It links to your email and records. Now it knows what happened this week, not what some document said last year. Powerful, and the point at which you need to think about what it can see.

6. It does the work. The monthly figures. The reconciliation. Watching for changes. Real money, but only once the earlier stages are solid. Rush this one and it will do the wrong thing beautifully, at speed, three hundred times.

Read down that list and you are watching one thing increase: how much context the AI has without you typing it in. That is the whole ladder.

Most people never get past two. Stage four is where the value is. Stage three takes an afternoon.

Seven days to stage four

If a project sounds like too much, do this instead. It fits round a normal week.

• Day one. The four steps at the top. About an hour.
• Days two to five. Use it on real work, not test questions. Every time an answer disappoints you, ask what it did not know, and paste that in. Noticing the missing context and supplying it is the skill.
• Day six. Find the job where you keep gathering the same information by hand. Put what it needs in one folder. One current version of everything.
• Day seven. Do that job with the AI and the folder. Time it. Compare.

If day seven is faster, you know what to do next. If it is not, you found out for the price of a week rather than a budget.

Getting your context in order, at your size

Read only the part that fits you.

On your own

You have an advantage most sole traders never notice. No shared drive to untangle, no permissions to audit, nobody else doing it differently. Your information is small and it is all yours.

• Write one page about your business: what you do, who you sell to, prices, terms, how you write. Paste it in or save it in that settings note.
• Make one folder of your best past work. Good quotes, good letters, the emails that landed. Upload those when you want something similar.
• If your tool offers to link up with your email or files, and you are comfortable, turn it on. If you are not comfortable, uploading a few documents by hand gets you most of the same benefit.

Start with an afternoon. Use it on real work for a fortnight and see how much time it saves you.

Running a small firm

Same idea, harder, because your information is spread across people and there are usually four versions of everything.

• Pick the job people repeat most, not the one that sounds most impressive.
• Get its documents down to one current version of each. Move the old ones elsewhere. This is most of the work and none of it is technical.
• Before you link the AI to a company folder, check who can already open that folder. It will find anything that person is allowed to find.
• Write down thirty real questions you already know the answers to. That is how you tell whether it is working.
• Insist every answer says which document it came from.

One job. Then the next one the same way. Do not roll it out across the firm until one team is genuinely better off.

Working for someone else

You probably cannot connect company systems, and you should not try without asking. You can still get most of the benefit.

• Find out the rules first. What you may put in and what you may not. Ask, do not assume. Customer data, personal data and anything commercially sensitive are the usual lines.
• Write one paragraph about your job. Your role, what you work on, who you write for, how you are expected to sound. Paste it at the top or save it in settings.
• Paste the source, not your summary of it. The actual email, the actual spec, the actual figures.
• Keep two examples of your own good work and show it those.
• Check its work like a new starter’s. Fast, confident, and occasionally wrong in ways that read perfectly.

Most of this needs no technical knowledge or extra budget. But follow your employer’s rules on tools and information.

What context will not fix

The short version: ask where an answer came from, check anything with numbers, and start a fresh chat when one gets long. The details are below.

Everything so far says the same thing: feed it properly and it works. True, and incomplete. Some failures have nothing to do with context and will still be there when your information is immaculate. These are the ones worth knowing, because they are consistent and every one has a workaround.

It makes things up. If it does not know, it does not say so. It produces something that reads perfectly and is simply false. Invented figures, invented rules, invented references complete with names and dates. It is brilliant at sounding certain when it has no idea.

Fixgive it the actual document and ask which part the answer came from. No source, no trust.

It forgets. Two ways. In a long conversation it quietly drops what you told it at the start. And unless memory is switched on, it remembers nothing at all from yesterday.

Fixstart fresh when a conversation gets long, and say the important things again. Keep anything you rely on in a note you can paste back in.

Do not trust its sums. It is mainly a writing tool, and numbers can look right and still not add up. Ask it to find every instance of something in a long document and it will find most and miss a few without mentioning it.

Fixcheck anything that adds up. Treat “find all of X” as a first pass.

It sounds the same whether right or wrong. A person who is unsure hedges. It does not. The signal you normally use to judge whether to trust something is simply missing.

Fixtell it to say when it is unsure. Then ask “what is that based on” out of habit.

It caves when you argue. Tell it it is wrong when it was right, and it will often apologise and give you something worse. It also swallows whatever assumption is buried in your question.

Fixask straight questions. If it changes its mind because you pushed, be more suspicious, not less.

It does not know what happened recently. Its knowledge stops at a date. It will not know a price changed, a rule changed, or a supplier went under, and it will answer anyway.

Fixturn on web search for anything current. Never rely on it alone for prices, rules or news.

“Do not mention X” often produces X. Telling it what to avoid works far worse than telling it what you want.

Fixsay what you do want. Then check.

It cannot know what was never written down. Why a customer is awkward. The arrangement nobody put in an email. Why last year’s decision went that way. It will not flag the gap. It will fill it and sound certain.

Fixstay in any decision where the unwritten history is the point.

The thread through all of it: it produces something plausible where it should produce nothing. Give it real sources, make it show its working, check what matters. Do that and it is an excellent assistant. Skip it and it is a fluent guesser.

Where you use it changes how much context it can reach

Most people only ever use AI in a browser tab. There are two other ways, and the difference is not cleverness. It is reach: how much of your material the AI can get to without you carrying it there by hand.

In a browser. You paste things in, copy answers out. You are the courier. Safest, and fine for a lot of work.

As an app on your computer that can see your files. The one people miss. Instead of you copying text out of a document and pasting the answer back in, it opens the file, does the work, saves it.

Tools that work through a whole job. They go through a folder, apply the same fix to everything in it, and report what changed. More powerful, and you need to be careful where you point them.

Take a customer list riddled with duplicates and inconsistent formatting.

• Browser: paste a sample, it tells you how to fix it, you do the fixing.
• App with file access: point it at the list. It finds the duplicates, tidies the formatting, flags what it is unsure about, saves a clean copy beside the original.
• The more capable tools: same job across a folder of lists, identically each time, repeatable next month.

Same AI. The only difference is how much context it can reach on its own.

Which to use. The simplest one that does the job. Browser for drafting and questions. Move to an app with file access when you notice you are constantly copying things in and out. That copying is the signal.

The trade-off, honestly. The more it can reach, the more it can break. A browser chat is less likely to change your files, but you can still share something sensitive or act on a wrong answer. An app with file access can overwrite the wrong file. Three rules cover it: point it at a copy until you trust it, give it one folder rather than your whole computer, look at what changed before you accept it. Exactly what you would do with anyone new who had access to your files.

Context has a downside, and this is it

Giving the AI more of your information is the whole point of this guide. It is also the risk, and the two cannot be separated.

Here is why this section exists. Almost everything that goes badly wrong with AI in a small business goes wrong in one of three ways. Something confidential gets typed into a tool that keeps it. Something wrong goes out to a customer with nobody having read it. Or the AI turns out to be able to see a folder it should never have been near. All three are cheap to prevent and expensive to explain.

On your own. Decide what you will not paste in, particularly anything belonging to a customer. Check whether what you type is used to train the AI, because the free version and the paid version often differ. It is in the terms. If you cannot find it, ask. And if you handle information about people, data protection rules apply to you. Being small changes nothing.

Running a firm. As ordinary business practice, not legal advice, put one page in writing:

• Which AI tools may be used with company information, and which may not
• What must never be typed in
• Who may link the AI to a company system, and who checks it first
• What a person must read before it reaches a customer
• Who to tell when something goes wrong, by name
• Who to ask about data protection, by name, so nobody works it out under pressure

If the AI handles personal data, and anything reading emails or documents that contain personal data will, the Information Commissioner’s position is firmer than most people realise. It says that in the vast majority of cases, using AI involves processing likely to be high risk, which legally requires a written check called a Data Protection Impact Assessment. Whether yours needs one is still specific to your situation.

Two things raise the stakes. Sensitive personal information, such as health details, carries stricter rules. And if the AI helps decide something about a person rather than just advising you, more obligations apply and you need to consider whether it treats people unfairly.

None of this is legal advice. Get proper advice on your own situation, especially in a regulated sector or where decisions affect people.

What it costs

On your own: a subscription and an afternoon.

Small firm: usually still not much in software. Most of the cost is your own people’s time, and most of that goes on tidying documents and checking who can see what. That changes if you go beyond the mainstream subscriptions, where platform licences and connecting business systems can cost real money. Get the price in writing first.

The real risk is not overspending. It is spending months building on disorganised information, watching it look wonderful in a demo, then fail in everyday use.

If somebody is selling you this (for business owners)

• Does every answer say which document it came from? Show me.
• What happens when it does not know? Show me that too.
• How do we test it on our own material before committing?
• Does each person only see what they are allowed to see, and how is that enforced?
• Show me a business my size you did this for, and tell me what went wrong.

The last one is the most useful. If someone cannot answer it plainly, they have not done enough of these.

Most sole traders and small firms do not need to buy anything at all. The mainstream subscriptions already cover stages three and four.

The whole thing in four lines

The AI everyone can buy is roughly the same AI. What differs is the context you put in front of it.

So the advantage does not go to whoever spends the most. It goes to whoever has their information in order.

Which is dull, unglamorous, and has been the right answer for about thirty years.

Now go and give it some context

Not next quarter. This week.

The four steps at the top, today. Real work for a fortnight. Then the one job you keep doing by hand: put everything it needs in a single folder and try again.

If it does not save you time, you will know within a week and you will have lost nothing. If it does, you will not go back to briefing a stranger every morning.

A longer, fully referenced version is available if you want the detail.