Sh*t in=sh*t out
Free guideHow to write a good prompt.
Most people get mediocre answers out of AI because they ask the way they would ask a search engine. This is the whole thing, from the simplest question to the hard ones.
What a prompt is.
A prompt is not a search query. It is a set of instructions to something that will do exactly what you asked, including the parts you did not mean.
That is the whole difference. Google guesses what you meant from three words. A language model does what you said. So if you leave out the format, the audience or the length, it picks for you, and it usually picks something long and generic.
The same question genuinely does give different answers each time. That is not a fault, it is how these systems work. A better prompt narrows the range so the answers you get are all useful ones.
The five things every good prompt has.
Get these five in and you are already ahead of most people using these tools every day.
-
What you want
Not the topic. The actual thing you want to end up holding. "A quote email" beats "help me with quoting".
-
Who it is for
A customer, your accountant, a new apprentice, or yourself. It changes every word of the answer.
-
What it already needs to know
Paste the actual details. The real email, the real numbers, the real job. Do not describe them.
-
What good looks like
Length, tone, format. "Four short paragraphs, friendly but not chummy, no bullet points."
-
What to leave out
The most underrated one. "Do not mention price." "Do not apologise." "Do not use the word partner."
Asking well.
For simple, everyday jobs. Write an email, tidy up some text, explain something.
Write an email to a customer about a delay.
You will get four paragraphs of corporate apology that sounds nothing like you, and you will rewrite all of it.
Write a short email to a customer named Sarah. Her job was booked for Thursday and the part has not arrived, so it moves to the following Tuesday. Three short paragraphs. Apologise once, clearly, then move on. Offer the new date. Do not offer a discount. Plain Australian English, the way a small business owner would write.
Everything it needed was in the prompt, so nothing had to be guessed.
- Say the format. Email, table, bullet list, one paragraph. If you do not, you get an essay.
- Say the length. "Under 120 words" works better than "short", which it will ignore.
- Give it the raw material. Paste the actual email you are replying to rather than describing it.
- Name the reader. Writing for a customer and writing for your bookkeeper are different jobs.
- Say what not to do. This does more work than almost anything else.
Show, do not tell.
The single highest value habit in this guide. If you want something to sound a particular way, paste an example of it rather than describing it.
Research on prompting keeps landing on the same finding: well chosen examples do more than extra instructions. An example shows the model what good looks like in a form it can copy directly, where an instruction has to be interpreted first.
The practical version for a business owner: stop trying to describe your tone of voice. Paste two emails you have sent and say "write the next one like these".
One warning that comes with it. Modern models pay very close attention to the details in your examples, including the ones you did not intend to teach. If both your examples happen to start with "Hi mate", every reply will start with "Hi mate" forever. Pick examples that are typical, not just handy.
Write it in a friendly, professional tone that matches our brand voice.
Means almost nothing. Every business thinks it is friendly and professional.
Here are two emails I have sent before. [paste them] Write the next one in the same voice, about a rescheduled job.
Nothing to interpret. It has the pattern in front of it.
Thinking step by step.
For anything with reasoning in it. Comparing options, checking numbers, working out which of three quotes is cheapest.
Asking a model to work through something in steps, rather than jumping straight to an answer, measurably improves it on problems that have reasoning in them. The phrase people use is chain of thought, and in practice you get it by writing something like "work through this step by step before you answer".
Two things worth knowing.
It is not free, and it is not always better. On simple tasks it makes answers longer and no more correct. Save it for problems where you would want to see somebody's working.
Ask for the working, then the short answer. The pattern that works best is to let it reason at length and then require a short conclusion in a fixed format. Something like: "Think it through first. Then give me your answer as three bullet points and one recommendation." You get the benefit of the reasoning without having to read all of it.
Newer reasoning models do a version of this on their own, so you may not need to ask. It still helps to say what the final answer should look like.
OCCD, for the hard ones.
When the job is big enough that a paragraph will not cut it, use a structure. Four parts, in this order.
Outcome
Name straight up what you want. First line, no preamble. Not the topic, the deliverable.
Context
What it needs to know about your situation. The business, the reader, the history, the constraints that are already true.
Constraints
What it must and must not do. Length, format, tone, things to avoid, things that are non negotiable.
Definition of done
How you will know the answer is right. The part almost nobody writes, and the one that stops you going round in circles.
Help me write a proposal for a commercial client who wants ongoing maintenance work.
You will get a generic template with square brackets in it, and spend an hour fixing it.
Outcome. A two page maintenance proposal I can send tomorrow, in my voice, ready except for the numbers.
Context. The client manages four commercial sites in Brisbane. We did one job for them in March and it went well. They asked for quarterly servicing across all four. They care about not having to chase anyone. Decision maker is the operations manager, not the owner.
Constraints. Under 800 words. No jargon. Do not include pricing, I will add it. Do not promise response times. Australian spelling. Confident, not salesy. Do not mention that we are a small business.
Definition of done. An operations manager could read it in three minutes and know exactly what they get, what we need from them, and what happens next. If any sentence could be in anyone else's proposal, rewrite it.
The last line does more work than everything above it. It gives the model a test to check its own answer against.
Why long conversations get worse.
This is the thing almost nobody tells business owners, and it explains most of the frustration people have with these tools.
A model does not read a long input the way you read a page. Researchers have measured what happens as you feed in more, and the answer is consistently that accuracy falls. Two findings matter in practice.
Where you put the important bit changes the answer
Accuracy when the key information sits at different positions in a long input. Highest at the start and the end, worst in the middle.
See the numbers
| Position | Accuracy |
|---|---|
| Very start | about 75% |
| Quarter of the way in | about 57% |
| Middle | about 52% |
| Three quarters in | about 56% |
| Very end | about 65% |
Shape redrawn approximately from Liu and others, Stanford, 2023, on what researchers call the lost in the middle effect. The same pattern was found across models from several different companies. Figures are illustrative rather than exact.
The more you paste in, the worse it does
Accuracy against how much material is in the input. Longer is not better.
See the numbers
| Amount of input | Accuracy |
|---|---|
| A few paragraphs | about 75% |
| Around ten documents | about 66% |
| Around twenty documents, roughly 4,000 words | about 57% |
Figures approximate, from the same body of research. Later work in 2025 found this is driven by the sheer length of the input rather than by the model failing to find things, and that once a context window is more than about half full, the most recent material wins and the earliest material fades.
What that means when you sit down to use it
- Start a new chat for a new job. Do not keep one long conversation going all week. It gets less reliable, not more informed.
- Put the important instruction at the top or the bottom. Never buried in the middle of a long paste.
- Paste only what it needs. More context is not better context. Twenty pages when three would do makes the answer worse.
- If answers start drifting, that is expected. Start again and paste in just the conclusion you had reached, not the whole history.
- Repeat the key instruction at the end of a long prompt. It costs you one line and it lands in the strongest position.
Why this happens, for the curious
A model weighs every piece of the input against every other piece. The further apart two pieces are, the weaker that connection gets, and the maths that normalises the result concentrates whatever attention is left onto the strongest matches. The beginning and the end of an input are structurally advantaged. The middle is where things go to be forgotten.
The ones almost everyone makes.
- Asking two things at once. "Write the email and also tell me if I should charge more." Split it. You will get a worse version of both.
- Being polite instead of clear. "Could you possibly help me maybe draft something" wastes the space where your instructions should be. It is not rude to be direct.
- Correcting instead of restarting. Three rounds of "no, not like that" leaves all three wrong versions sitting in the conversation, pulling the next answer toward them. Start again with a better prompt.
- Assuming it remembers. Unless you have turned on a memory feature, a new chat knows nothing about you, your business or last Tuesday.
- Believing a confident number. These systems state wrong figures in exactly the same tone as right ones. Check anything that matters.
- Not saying the format. The single most common reason an answer is unusable.
- Pasting confidential material without checking where it goes. Know your provider's settings before customer data goes anywhere near it.
Templates for the usual jobs.
Replace the square brackets. That is the whole job.
Replying to a difficult email
Here is an email I received. [paste it] Write my reply. I want to [what you want to happen]. Keep it to [length]. Firm but not cold. Do not apologise for anything that was not our fault, and do not offer [thing you are not offering]. Australian English.
Summarising a long document
Here is a document. [paste it] Summarise it for [who]. I need to know [the specific thing you need to decide]. Give me five bullet points, then one line on what I should do about it. Ignore anything not relevant to that decision.
Drafting a quote or proposal
Outcome: a [document] I can send today, in my voice. Context: the client is [who], they want [what], and they care most about [what]. We have [history]. Constraints: under [length]. No pricing, I will add it. No jargon. Do not promise [things you cannot promise]. Definition of done: [who] could read it in [time] and know exactly what they get and what happens next.
Cleaning up a spreadsheet
Here is some messy data. [paste it] Put it into a table with these columns: [list them]. Where a value is missing, write "unknown" rather than guessing. Do not change any numbers. Tell me at the end which rows you were not sure about.
Writing a job ad
Write a job ad for [role] at a [size] [type] business in [location]. The person will mostly [main tasks]. They need [must haves]. We can offer [what you offer]. Under [length]. Written the way a person talks, not a recruitment agency. No "rockstar", no "fast paced environment", no "wear many hats".
Making sense of your numbers
Here are some figures from my business. [paste them] Work through this step by step before answering. I want to understand [the actual question]. Show me the working, then give me three bullet points and one recommendation. If the data is not enough to answer properly, say so rather than estimating.
Rather it just happened without you?
Writing good prompts is worth learning. Not having to write them at all is better. That is roughly what we build.
Book a call