Practical guide
You've seen the two modes.
Here's what to do with them.
The theory is simple. The practice takes a little more.
The framework
Think. Collab. Dispatch. Iterate.
Four steps. Most people skip straight to the third one. That's why it's frustrating.
Step 01
Think
Get your messy human thoughts together before you open the machine. Brain dump, voice memo, back of an envelope — anything. The machine is brilliant at structuring chaos. But it needs your chaos first.
Don't skip this. It's the one that changes the quality of everything after it.Step 02
Collab
Use the machine to figure out what you actually want. Ask it questions. Ask it to ask you questions. Let it sharpen the brief before you write the brief. You don't have to know exactly what you want going in.
The most underused step. The one that makes Dispatch fast.Step 03
Dispatch
Write a clean instruction and send it. You don't need to watch it work. Define the output, the format, the length, the audience — then walk away. The machine doesn't need supervision.
Dispatch and disappear. Check back when it's done.Step 04
Iterate
One or two rounds, then stop. Quality degrades with over-iteration — every round takes you further from your original intent. Define what done looks like before you start, and stop when you're there.
Infinite iteration is a losing game. The first good version is usually the best.The machine doesn't need breaks. You do. Walk, eat, talk to people. The best thinking happens away from the screen — that's where the real ideas come from. Take rest seriously. The work will be there when you get back. So will the machine.
Same task. Different input. Very different output.
What good input
actually looks like.
The better versions are still plain English — just specific. You don't need to write like a robot. You just need to stop being vague.
The vague ask
Less effective
"Write me something about AI for my blog."
more effective
"Write a 350-word blog intro for small business owners who've never really used AI. Hook it with a specific scenario — someone trying to get ChatGPT to write their emails and feeling more confused than before. Conversational and honest, no jargon. End with a question that makes them think."
Still plain English — just specific. The first version leaves every decision to the machine. The second one makes the decisions you can make, so the machine can focus on the ones it's good at.
The apology reflex
Less effective
"Hi! Hope you're having a good one 😊 So sorry to bother you — I know this is probably a silly question, but I was just wondering if you could maybe explain the difference between machine learning and AI? Only if you have time, absolutely no rush!"
more effective
"What's the difference between machine learning and AI? Keep it simple — I don't have a technical background. Two paragraphs."
Nobody has ever hurt a chatbot's feelings. The greeting, the apology, the hedge — skip all of it. The question is the same. The output is the same. The only difference is the energy you spent getting there.
Everything at once
Less effective
"Can you write a follow-up email, summarise the meeting notes, create an action list with owners, suggest a subject line, and check if the tone is right for the situation?"
more effective — one at a time
Scope creep in the prompt equals scope creep in the output. One thing at a time produces better results and makes it easier to course-correct.
No context vs. the right context
Less effective
"Write a proposal."
more effective
"Write a one-page project proposal. It's for a Dublin restaurant — we're redesigning their online ordering system. Budget is around €12k, eight weeks. My company builds web apps for small businesses. Professional and direct, no waffle."
Context is a precision tool — give it what it needs to do the job well, nothing more. The difference between "write a proposal" and this version isn't syntax. It's specificity.
No role vs. giving it a role
Less effective
"Give me feedback on this presentation deck."
more effective
"You're a senior consultant who's pitched to hundreds of executives. Look at this deck and tell me the top three things that'll lose a C-suite audience. Be direct — I can handle it. Worst issue first."
"You are a..." is the single highest-leverage thing you can add to any prompt. It calibrates tone, assumptions, and quality all at once — and it takes five seconds.
Endless iteration vs. defining done
Less effective
"Make it better." → "More engaging?" → "Actually more formal." → "Try a completely different approach." → "I liked the first version, can you go back to something like that but with a different ending?"
more effective — define done first
"Write a product description for [X]. It's done when it's under 150 words, has one clear benefit per sentence, no superlatives, and ends with a specific call to action. Tell me when you think it's there."
Define what done looks like before you start. The machine will keep going until you tell it to stop — "make it better" has no end state. "Done means X" does.
Working with machines
Eight things worth knowing.
Not rules. Just patterns that consistently make things better.
01
Skip the greeting.
Start with the task.
The preamble costs you energy and changes nothing. "Hi! Hope you're well!" to a language model is energy spent on something that cannot receive it. Start with what you need.
02
Give it a role
before you give it a task.
"You are a senior editor who has worked on 500 articles" changes the output more than almost anything else you can add. It calibrates tone, assumptions, and quality all at once.
03
One thing at a time.
Every time.
Scope creep in the prompt equals scope creep in the output. Ask for one thing, get a result, review it, then ask for the next. The quality of everything improves when you do this.
04
Context is a precision tool.
Use it like one.
Give it what it needs to do the job well — nothing more. Too little context produces vague output. Too much and it gets lost. Think: what does it actually need to know here?
05
Define done
before you start.
"Done means: under 200 words, one call to action, no jargon." Setting the success criteria upfront stops infinite iteration before it starts. The machine will keep going until you tell it to stop.
06
Done is done.
Stop at the first good version.
Quality degrades with over-iteration. Every round takes you further from your original intent. The first good output is usually the best. Recognise it and stop.
07
If it's going wrong,
start a new conversation.
Don't fight a bad thread. The machine carries its context forward — if it's going sideways, you're fighting that momentum with every message. A fresh conversation with a better brief is faster every time. Context creep is real.
08
Talk to it directly.
Ask it what it needs.
"What information would help you give me a better answer?" is a completely valid question. Ask it to ask you questions. Ask it to interview you. Use the machine to build the brief that you then give the machine. It works.
One more thing
Don't be afraid
to have a conversation with it.
You don't have to arrive at the machine with a perfect brief. You can use the machine to build the brief. Ask it to ask you questions. Tell it what you're trying to do and ask what it needs to know. Let it interview you.
That conversation produces a brief. The brief produces good output. The whole process takes less time than staring at a blank prompt wondering where to start.
This is how you build a working pattern, not just a one-off interaction. The machine will tell you exactly what it needs. Listen to it. Save what works.
See the difference
for yourself.
Feels different when you try it. No login. No sign-up.