Most people use Chat GPT 4 like a search box with better manners. They type a question, get three paragraphs, copy them into a document, and move on. That works fine for trivia. It falls apart the moment the output has to be correct, on-brand, and finished before Thursday.
The gap between a frustrating session and a useful one is mostly process, not prompt magic. Here is the six-step routine I run for anything that matters, with the actual wording I type.
Step 1: Name the role, the reader, and the shape of the output
Vague instructions get vague results. Compare these two openers.
Weak: “Write an email about our price increase.”
Better: “You are a customer success manager at a 20-person accounting firm. Write an email to long-standing clients explaining a 7% fee increase from 1 March. Tone: matter-of-fact, no grovelling. Maximum 150 words. Include one sentence on what stays the same.”
The second version takes forty extra seconds and removes about three rounds of correction. If the setup stage feels awkward, this guide to briefing an AI chat like a new colleague covers it properly. The mindset shift is treating the model as a capable freelancer on their first day, not a mind-reader.
Step 2: Paste the raw material instead of describing it
Numbers, quotes and constraints belong in the prompt. If you want a summary of customer feedback, don’t write “our customers are unhappy about delivery times.” Paste the twenty actual complaints. Chat GPT 4 handles a few thousand words of pasted context without complaint, and the specificity of the source material shows up in the output.
Style works the same way. Paste a paragraph of your own writing, say “match this voice,” and it will.
Step 3: Ask for three options at different levels of ambition
One answer invites yes-or-no thinking. Three answers give you a real choice. A prompt that works well:
- Option A: safe and formal, suitable for a board update.
- Option B: conversational, suitable for a newsletter.
- Option C: the version you’d write if you weren’t worried about offending anyone.
You’ll usually take a sentence from A, the opening from B, and one idea from C. That mix is where the value sits. A broader five-step workflow for using an AI chatbot online describes the same diverge-then-converge rhythm in more general terms.
Step 4: Make it argue with itself
Once you have something decent, ask: “What are the three strongest objections to this plan?” Then: “Which of those is most likely to actually happen, and what would you change?”
This is where Chat GPT 4 earns its keep. It tends to be better at spotting a gap in your reasoning than at generating a first draft from nothing. Run the same trick on emails, budgets, project plans and itineraries. If you want to see it applied across an entire project, there’s a step-by-step walkthrough of planning an event with a ChatGPT chatbot that shows the back-and-forth in detail.
Step 5: Tighten with constraints, not adjectives
“Make it better” achieves very little. “Cut it to 120 words, delete every adjective, and rewrite the opening sentence so it starts with a verb” achieves a lot. Constraints are measurable; adjectives are not.
A short list worth keeping nearby:
- Half the length, same meaning.
- No sentence over 20 words.
- Swap every abstract noun for a concrete one.
- Rewrite for a reader who has 15 seconds.
- Delete the last paragraph. It’s usually a summary nobody asked for.
If the output is heading anywhere near a customer’s inbox, an auto-reply, or a help-centre answer, one pass is never enough. There’s a separate playbook on building a chatbot AI that survives real customers worth reading before you put generated text in front of paying people.
Step 6: Verify anything that has to be true
Chat GPT 4 sounds equally confident whether it’s right or wrong. Statistics, dates, legal requirements, pricing bands and product specs all need a second source. Numbers it produces for “market size” or “average conversion rate” are frequently invented, several years out of date, or lifted from a US study that doesn’t apply to your market.
The practical rule: let the model do the drafting, structuring and editing, then do the fact-checking yourself. The real-world capabilities and limitations of Chat GPT 4 are well documented by now, which means you can predict which task types will need extra scrutiny before you even open the chat window.
A worked example: turning a 45-minute interview into a case study
A two-person bookkeeping firm recorded a 45-minute call with a happy client. Here’s how the six steps play out in roughly 25 minutes of actual work.
Setup
Role: B2B copywriter. Reader: owners of £500k–£2m turnover businesses who currently do their own books badly. Output: a 600-word case study with three subheadings and one pull quote.
Input
The full transcript, pasted, plus two existing case studies so the voice matched.
First pass
Ask for three possible structures: chronological, problem-first, and results-first. Results-first wins for this audience almost every time.
Pressure test
“What is missing that a sceptical reader would want? List five gaps.” It came back with: the client’s previous system, the cost of switching, how long onboarding took, what happened at year-end, and whether the client would recommend them. Four of those five weren’t in the transcript, so the firm had to phone the client back and ask.
Compression
The draft landed at 940 words. One instruction, “cut to 620 words without losing any numbers,” did most of the heavy lifting.
Verification
Every figure in the final piece came from the transcript or the firm’s own records. Nothing was taken on trust from the model.
Total time: about 25 minutes including the follow-up call. Done manually, the same piece would have eaten half a day.
The habit that makes the difference
Keep a running document of prompts that produced something you actually used. After a month you’ll have twelve or fifteen of them, and most of your regular work will start from a saved opening rather than a blank box. The models will keep changing names and numbers. The loop holds up: brief, feed, branch, pressure test, constrain, verify.
One thing worth internalising, especially if you’re new to this: the first response is rarely the answer. It’s the material you refine. Treat it that way and Chat GPT 4 stops feeling like a novelty and starts behaving like a genuinely useful member of the team.