How to Talk to AI
Five practical techniques that turn vague AI output into something you can actually use.
The way you phrase your instructions to AI is called prompting, and it matters more than most people realise. The quality of the output is entirely dependent on the quality of the input. This is not a slight difference. It is the difference between getting back something generic and useless versus something that saves you hours of work. Let me show you exactly what I mean.
The Lazy Prompt vs The Crafted Prompt
Here is the same task, done two ways.
The lazy prompt:
You will get back a generic template with placeholder text. It will be bland, formulaic, and require complete rewriting. You have wasted both your time and the AI's processing power.
The crafted prompt:
My company: Assure Advantage, a specialist bid assurance consultancy. We help organisations identify structural risks in large IT procurement before they commit to contracts.
The client: A mid-tier Australian government department about to go to market for a core systems replacement worth approximately $50-80 million over 5 years. They have had two failed IT procurements in the past decade.
What I am proposing: An independent structural risk assessment of their draft RFP and evaluation framework before they release it to market. This is a 4-6 week engagement.
What I need: A 2-page proposal (not longer) that opens with their problem (failed procurements and the cost of getting this one wrong), positions our approach as different from standard risk consulting (we assess the structural design of the deal, not just the risk register), outlines the scope and timeline, and states the investment.
Tone: Authoritative but not arrogant. Peer-to-peer, as if written by someone who has seen this problem many times and knows how to solve it. No consulting jargon. No buzzwords. Short sentences. Direct language.
Do not include: Executive summary headings, "about us" sections, or any boilerplate. Start with their problem, not with us."
The second prompt gives the AI everything it needs: who you are, who the client is, what the situation is, what you want delivered, how long it should be, what tone to use, and what to avoid. The output from this prompt will be something you can actually send with minor editing, not something you need to rewrite from scratch.
That is the core lesson of prompting. The more context, specificity, and constraint you provide, the better the result. Every detail you include eliminates a decision the AI would otherwise make on its own, usually badly.
The Anatomy of a Good Prompt
A well-crafted prompt typically includes some or all of these elements:
Role: Who should the AI be? "You are a senior HR consultant with 20 years of experience in Australian employment law." This frames the vocabulary, the depth, and the perspective of the response.
Context: What is the situation? The more background you give, the more relevant the output. Do not assume the AI knows anything about your specific situation.
Task: What exactly do you want? "Write," "Analyse," "Compare," "Identify," "Draft," "Critique." Be precise about the action.
Constraints: What are the boundaries? Length, format, tone, audience, things to include, things to avoid. Constraints are where good prompts become great prompts.
Examples: If you want output in a specific style or format, show the AI an example. "Here is an email I wrote last week that I was happy with. Match this tone and style." This is one of the most powerful prompting techniques and the most underused.
Output format: Tell it exactly how you want the response structured. "Give me a table with four columns." "Structure your response as: first the diagnosis, then three options, then your recommendation with reasoning." "Use short paragraphs, no bullet points, no headings."
Example: Preparing for a Difficult Conversation
Here is a long prompt that demonstrates how much context changes the output:
I am a senior project manager in an IT services company. I have a team member, let us call him James, who has been with the company for 8 years. He is technically excellent but increasingly difficult to work with. He pushes back on every process change, openly criticises decisions in team meetings, and has started being dismissive toward two junior team members who joined in the last 6 months.
I have had two informal conversations with him about the behaviour. Both times he acknowledged the issue but nothing changed. My manager has asked me to have a formal conversation and document it.
I genuinely respect James's technical ability and do not want to lose him. But the impact on the team is real and getting worse.
Help me prepare for this conversation. I need:
1. An opening statement that is direct but not aggressive. I want to be clear this is formal without making it feel like an ambush.
2. Three specific examples I should raise (suggest the kinds of examples based on the behaviours I described, since I will substitute my own real ones).
3. How to handle the most likely pushback: that he will say the junior staff are not good enough and that processes are bureaucratic.
4. A way to end the conversation that gives him a clear path forward rather than just a reprimand.
5. What to document after the meeting.
Context: We are in Australia. I need to be mindful of fair work obligations. The company has an HR team but I want to handle this at my level first if possible."
That prompt gives the AI a specific person, a specific situation, a history of prior attempts, your emotional state (respect for the person, frustration with the behaviour), exactly what outputs you need, anticipated obstacles, and jurisdictional context. The response will be practical, specific, and immediately usable. Compare this to "Help me deal with a difficult team member" and you can see why prompting quality matters.
Example: Analysing a Complex Document
When you upload a document to AI, do not just say "summarise this." Tell it what you need and why:
I need you to analyse this document and give me:
1. A one-paragraph summary of what they are actually buying (not their official description, but what the scope really amounts to in practical terms).
2. The five most important evaluation criteria and their weightings. For each one, tell me whether it favours incumbents, specialists, or large generalists.
3. Any unusual or onerous contract terms that might make this deal unattractive.
4. Red flags: anything in the document that suggests the outcome may already be predetermined, the timeline is unrealistic, or the budget is insufficient for what they are asking.
5. Your honest assessment: is this worth bidding on for a mid-sized IT services company with strong infrastructure capability but no existing relationship with this department?
Be direct. I would rather hear 'do not bid, here is why' than a balanced both-sides analysis that does not help me make a decision."
This prompt turns AI from a summariser into a strategic advisor. You have told it your situation, what you need to decide, how to structure the analysis, and that you want a direct recommendation rather than hedge-everything waffle. The output will be something you can take into a bid/no-bid meeting.
Example: Creating Content With a Specific Voice
One of the hardest things for AI is matching a specific writing voice. The solution is to give it a sample and be explicit about what makes that voice distinctive:
[paste 500-1000 words of your own writing]
Before you write the article, tell me what you observe about my writing style: sentence length, tone, use of examples, level of formality, how I handle technical content, and any distinctive patterns. I want to verify that you have understood my voice before you attempt to replicate it.
Once I confirm, write a 600-word article about why most IT risk registers fail to identify structural problems in large deals. The article should follow the same structure as my sample: open with a specific observation, explain why the conventional approach does not work, describe what should be done instead, and close with a direct challenge to the reader."
The key technique here is asking the AI to describe your voice back to you before it writes. This forces it to explicitly identify the patterns rather than making vague approximations. If its description does not match what you think your voice is, correct it before it writes. This two-step process (analyse then write) produces dramatically better results than "write something in my style."
Iteration: The Real Skill
Your first prompt rarely produces a perfect result. That is normal. The real skill is in the follow-up. Good iterative prompts are specific about what is wrong and what you want changed:
"The second paragraph is too vague. Replace the general claim with a specific example from Australian government IT procurement. Reference a real, publicly known project failure."
"The tone has drifted into consulting-speak. Rewrite the whole piece as if you are explaining this to a smart colleague over coffee, not presenting to a board."
"This is too long. Cut it to 400 words. Keep paragraphs 2 and 5 intact. Condense everything else."
"You have given me a balanced analysis. I asked for a recommendation. Pick a side and defend it."
"The structure is right but the opening is weak. Give me three alternative opening sentences that start with a provocative statement or a specific fact, not a general observation."
Three to five rounds of focused iteration typically produces something excellent. The people who get disappointing results from AI are almost always the ones who accept the first output or give vague feedback like "make it better." Be as specific in your feedback as you were in your original prompt.
Business Prompts (With Privacy Warnings)
AI can be enormously useful for business tasks, but you need to understand what you are sharing and with whom.
Competitive analysis: "Research [competitor name] and summarise their service offerings, pricing model, and market positioning. I want to understand where they are strong and where there might be gaps we could exploit." This is safe because you are asking about publicly available information.
Document review: "Review this contract and highlight any clauses that are unusual, one-sided, or that I should negotiate." Privacy warning: You are uploading a potentially confidential document. See the privacy section before doing this with sensitive contracts.
Financial analysis: "Analyse this spreadsheet and identify the top 3 cost categories that have increased by more than 10% year-on-year. Suggest areas to investigate for cost reduction." Privacy warning: Financial data is sensitive. Use a paid tier (which offers stronger data protection) and never upload data with client names or personal information unless your privacy policy allows it.
Strategy brainstorming: "Act as a business strategy consultant. Our company provides [service] to [market]. Revenue has plateaued at [amount]. Give me 5 strategic options for growth, with pros and cons of each." This is generally safe as you are sharing business context, not personal data.