Prompt Engineering for Business — How to Talk to AI So It Actually Delivers

Imagine you hire an assistant. On their first day they walk into the office, sit at the desk and wait. You lean out from behind your monitor and say: "offer for client." Then you go back to work.

The assistant — even a very talented one — will write something generic, bland and useless. Not because they're incompetent. Because they don't know who the client is, what you want to achieve, what language your company speaks or exactly what you expect as output.

That's exactly what 90% of businesses do with their AI.

They type a keyword. They get garbage. They complain AI doesn't work. Meanwhile the problem isn't the model — it's the brief.

Prompt engineering is simply the skill of giving good briefs. Nothing more. You don't need a PhD or programming skills. You need to know what you want, who you're asking, and what the output should look like.

Diagnosis: Why Your Prompts Give Poor Results

Before I show you how to write well, let me show you why it goes wrong. After years of building AI systems and configuring prompts for clients, I've identified four patterns of bad briefing:

Pattern 1 — "The Search Engine" You treat AI like Google. You type three keywords and expect a precise result. "B2B sales proposal." "Email to client." "Competitor analysis."

AI doesn't read minds. It generates the most probable text matching your words — and that will always be something generic, because general phrases have generic answers in the training data.

Pattern 2 — "No Role" You don't tell the AI who it should be. By default, the model behaves like a helpful general-purpose assistant trying to be safe, neutral and inoffensive. For sales content, that's lethal.

Pattern 3 — "No Context" You don't tell the AI about yourself, the client, or the conversation context. The model answers an abstract question, not a specific situation. The difference is enormous — like the difference between a general medical article and a response from a doctor who knows your history.

Pattern 4 — "Any Format" You don't say what the output should look like. So you get an essay in three paragraphs instead of bullet points, a treatise instead of a table, a five-page plan instead of a one-sentence decision.

Four Components of a Good Brief — The RCST Framework

Through years of trial and error across dozens of deployments, one framework crystallised that works regardless of AI model or use case. I call it RCST — Role, Context, Style, Template (Format).

R — Role Who is the AI in this conversation? Sales specialist? E-commerce copywriter? Financial analyst? Legal assistant? The more precisely you define the role, the more specialist the response style. Role isn't just a job title — it's a point of view, the technicality level of language, and priorities in argument selection.

Weak: *"Write an email to a client."* Better: *"You are an experienced B2B account manager at a manufacturing company. Your communication style is direct, concrete, no corporate jargon."*

C — Context What does the AI need to know to do the job well? Who is the recipient? What's the goal? What information is available? What has already happened? The more context, the more precise the response. Don't worry about too much — models handle long context better than its absence.

Weak: *"Write a response to the complaint."* Better: *"Client John Smith filed complaint #2847 — product arrived damaged. He's been our client for 3 years. His order was delayed once last year too. We want to retain the client and offer product replacement plus 15% discount on the next order."*

S — Style How should the response sound? Formal or casual? Empathetic or matter-of-fact? Persuasive or informational? You can also provide style examples — a quote from one of your own emails you like stylistically. AI calibrates style well on examples.

Weak: *no mention of style* Better: *"Write in this style: professional, warm, no corporate jargon. Avoid: 'per your request', 'please be advised', 'kind regards'. Use active sentences."*

T — Template (Format) What should the output look like? Bullet list? Table? Email ready to send? JSON? Three sentences? Two versions to choose from? If you don't say — the model decides for you, and the decision rarely suits you.

Weak: *no mention of format* Better: *"Return: email subject in quotes, email body ready to paste, and one sentence recommending the best send time."*

Before and After — Four Business Scenarios

This is the heart of this article. Instead of theory — concrete transformations.

Scenario 1: Email to a Cold Lead

Bad prompt:

bad-prompt-lead.txt
Write an email to a potential client.

Good prompt:

good-prompt-lead.txt
### ROLE
You are a Senior Account Manager specialising in B2B sales to manufacturing companies.
Your style: direct, concrete, no "watered-down" language. Zero phrases like "I have the pleasure of".

CONTEXT - Company: Steelworks Ltd, steel component manufacturer, 80 employees - Contact: Peter Wilson, Operations Director (LinkedIn) - My product: AI Vision-based quality control automation system - What I know about the company: recently expanded their production line (LinkedIn info) - Email goal: schedule a 20-minute call, NOT sell

STYLE AND TONE Write peer-to-peer — manager to manager. One concrete question at the end. Avoid: superlatives, words like "innovative/leading/comprehensive", generic promises.

FORMAT Email subject (max 60 chars) + email body (max 120 words) + PS with social proof.

---

Scenario 2: Analysing an Inbound Client Email

Bad prompt:

bad-prompt-analysis.txt
Analyse this email and respond.
[email content]

Good prompt:

good-prompt-analysis.txt
### ROLE
You are a B2B client relations expert with 10 years experience in the IT industry.

TASK Analyse the client email below and return a structured analysis.

CLIENT EMAIL [paste email content]

REQUIRED OUTPUT (JSON) { "sentiment": "positive|neutral|negative|mixed", "main_intent": "one sentence", "hidden_concerns": ["list", "of", "reading", "between", "lines"], "response_priority": "high|medium|low", "recommended_action": "what to do within 24h", "response_tone": "empathetic|factual|apologetic|sales" }

---

Scenario 3: E-commerce Product Description

Bad prompt:

bad-prompt-product.txt
Write a product description for an office chair.

Good prompt:

good-prompt-product.txt
### ROLE
You are a copywriter specialising in B2C e-commerce. You understand the psychology of online purchases.

PRODUCT Name: ErgoMax Pro Chair Technical specs: height adjustment 42–54cm, 4D armrests, breathable mesh, 150kg capacity, 5-year warranty, price £649

BUYER Office worker aged 30–45, spends 8h at a desk, has back pain, buying for themselves or the company. Compares offers, price-sensitive, but will pay more for health and quality.

STYLE Benefits language (not features). Avoid technical jargon. Short, concrete sentences. Zero: "unmatched", "highest quality", "perfect".

FORMAT - Headline: 1 sentence (max 10 words), focused on the main benefit - Lead: 2–3 sentences (what it changes in the buyer's life) - 5 bullet points with benefits (formula: WHAT → WHY IT MATTERS) - Closing sentence with call to action Total: max 180 words

---

Scenario 4: Weekly Management Report

Bad prompt:

bad-prompt-report.txt
Write a weekly report for management based on this data.
[data]

Good prompt:

good-prompt-report.txt
### ROLE
You are a business analyst writing reports for an impatient board.
The board wants: decisions, not descriptions. Exceptions, not averages. Actions, not history.

INPUT DATA [paste data: sales, KPIs, tickets, whatever you have]

CONTEXT - Week 21/2026 - Previous week was -12% vs plan - Q2 summer campaign is running

OUTPUT FORMAT 1. SIGNAL (1 sentence): whether the week was good/bad and why 2. TOP 3 RESULTS (table: metric | value | change vs previous week) 3. TOP 2 PROBLEMS (what's happening and who owns it) 4. DECISIONS REQUIRED (list, max 3 items with deadline) No preamble, no summary at the end. The board has 2 minutes.

Advanced Techniques — When One Prompt Isn't Enough

For simple tasks one good prompt is enough. For complex ones — you need a chain.

Chain-of-Thought: Make AI Think Out Loud

Adding the phrase "Before answering, write out your reasoning step by step" dramatically improves the quality of analytical responses. A model that thinks out loud makes fewer logical errors than one that answers directly. Especially effective for contract analysis, calculations, risk assessment.

Difference in practice: - Without CoT: "Should I accept the client's offer?" → generic answer. - With CoT: "Analyse the offer. First list the contract's strong points, then the risks, then the recommendation." → specific and justified answer.

Few-Shot: Teach by Example, Not by Rules

Instead of describing the desired style — show two or three examples of what you want to receive. For the model, one good example is worth ten pages of description.

few-shot-classification.txt
Classify customer submissions into one category: COMPLAINT / QUERY / ORDER / CANCELLATION.

Examples: Input: "I ordered 3 weeks ago and still nothing has arrived" Output: COMPLAINT

Input: "Do you have this model in black?" Output: QUERY

Input: "I'd like to order 50 units of model XL-200" Output: ORDER

Now classify: Input: "{{submission_content}}" Output:

Iterative Refinement: Don't Rewrite — Converse

When the result isn't ideal, most people write a whole new prompt from scratch. Better: keep the conversation going. Stay in the same thread and say specifically what's not working.

Instead of: "Rewrite this but better." Write: "The second paragraph is too generic. Replace it with a concrete example from manufacturing that shows time savings. Keep everything else."

One targeted comment fixes more than restarting the entire prompt.

Templates to Take — Five Ready-Made Prompts for SMBs

Copy these and adapt to your company. Change names, industry data and format requirements — the rest works.

template-sales-proposal.txt
### ROLE
You are an experienced B2B salesperson in [YOUR INDUSTRY]. You write proposals that sell — not inform.

CLIENT CONTEXT Company: {{company_name}} Industry: {{industry}} Client problem: {{problem_description_from_CRM}} Approximate budget: {{budget}} Decision maker: {{job_title}}

OUR PROPOSAL {{brief_solution_description}} Price: {{price}}

TASK Write a sales proposal that: 1. Opens with the client's problem (not with us) 2. Presents the solution in benefits language 3. Includes one concrete numerical result from a similar client 4. Ends with a clear Next Step with a deadline

Length: max 300 words. No attachments, ready to paste into an email.

template-follow-up.txt
### ROLE
You are an account manager who just returned from a difficult sales meeting.

MEETING CONTEXT Date: {{date}} Client: {{first_last_name}}, {{job_title}} at {{company}} Topics discussed: {{topics}} Main objection: {{objection}} Agreed Next Step: {{next_step}} Deadline: {{deadline}}

TASK Write a follow-up email that: - Confirms the agreements without sounding like meeting minutes - Addresses the main objection with one concrete argument - Clearly states what you do, what the client does, and by when - Tone: professional, not corporate

Length: max 150 words.

template-complaint-response.txt
### ROLE
You are a customer service manager. Priority: retain the client, not win the argument.

COMPLAINT CONTEXT Client: {{client_data}} History: {{loyal_client_since_when}} Problem: {{complaint_description}} Our position: {{whether_we_are_at_fault}} Possible compensation: {{what_we_can_offer}}

RULES - Start by acknowledging frustration, not explaining yourself - Don't use: "as per our terms", "company policy states" - Give a specific resolution deadline - If we're at fault — admit it directly

FORMAT Email subject + body ready to send (max 200 words).

How I Test and Calibrate Prompts

There's no good prompt without testing. Here's my process:

Minimum 10 different inputs. One test proves nothing. A prompt that works for one client email may fail on the tenth because it has a different tone or unusual formatting.

Test edge cases. What if the input is empty? What if the client is aggressive? What if the text is in a different language than expected? A good prompt handles exceptions without extra instructions.

Measure concrete metrics, not "I like it". For classification prompts — accuracy on a hundred examples. For generative prompts — does the output work without editing in more than 70% of cases? If not, the prompt needs work.

Version prompts like code. Every change is a new version. Record what you changed and what the effect was. After a week of iteration without versioning you no longer know what was working.

Don't change two things at once. If you edit the role and the format simultaneously, you don't know what improved the result. Change one element, test, record the result, move to the next.

Eight Rules That Changed My Deployments

Instead of FAQ — rules I apply to every project:

1. Context is a more valuable resource than the model. GPT-3.5 with perfect context will beat GPT-4o with zero context on business tasks. Before you spend more on a model — invest in the quality of input data.

2. Role changes everything. The same prompt with "You are an assistant" vs "You are a sales director with 15 years of practice" gives fundamentally different answers. Test roles aggressively.

3. Negative constraints work. "Don't use the word innovative. Don't start with 'Of course'. Don't write more than 100 words." A list of prohibitions is often more important than a list of instructions.

4. Examples speak louder than instructions. If you have three examples of good output from earlier work — paste them. The model will learn more from examples than from five paragraphs of style description.

5. Format is half the battle. Defining the output structure — JSON, table, list, headers — eliminates half the "do it again but differently" iterations.

6. A production prompt is not a test prompt. In testing you experiment. In production you have a locked prompt that has passed calibration and doesn't change without a testing process. Never edit a production prompt on a live system.

7. Temperature matters. For analytical and classification tasks — temperature 0 or 0.1 (deterministic). For creative tasks — 0.7–0.9. For extracting data from structure (JSON extraction) — always 0.

8. Prompt length is not a problem. Models today have context windows from 32k to 200k tokens. A 500-word prompt costs nothing. A good, long prompt beats a short sloppy one every time.

---

If you're building AI automation at your company and want to be certain the prompts in your system work reliably — get in touch. Prompt calibration for production systems is a separate service I offer with every deployment. A good prompt is the difference between a system that works 70% of the time and one that works 95%.