There’s a specific moment when ChatGPT stops feeling like a party trick. For a lot of people it arrives late at night, staring at a 3,000-word interview transcript and a deadline, wondering whether pasting the whole mess in will actually help. It usually does. The trick isn’t the prompt. It’s the habits around the prompt.
Most people never get that far. They ask one vague question, get one vague answer, and quietly file the tool under overhyped. The gap between those two experiences has nothing to do with technical skill. It’s about twenty minutes of practice.
Give ChatGPT a job before you give it a question
Default ChatGPT is agreeable. It will answer almost anything competently and finish very little of it well. The fastest fix is to tell it who it is before you tell it what you want. Two openers for the same task:
- Weak: “Write a job ad for a barista.”
- Stronger: “You run a busy independent coffee shop in Leeds. You’ve read 200 barista applications this month and most were obvious copy-paste jobs. Write a job ad that makes a genuinely good barista stop scrolling.”
The second version carries the setting, the frustration and the competition. ChatGPT only knows what you type, so every detail you leave out is a detail it invents.
Skip the flattery
“Act as a world-class expert with 20 years of experience” adds nothing except words. What changes the output is the constraint: the audience, the length, the tone, and the thing you specifically don’t want. A role like “you’re a hiring manager who hates jargon” beats “you’re a genius recruiter” every single time.
Paste the raw material, not your summary of it
This is the habit that separates people who find ChatGPT useful from people who find it forgettable. Instead of describing the problem, hand over the actual thing.
- Don’t write “the client wants the tone warmer.” Paste their last three emails.
- Don’t type “here’s roughly what we discussed.” Paste the meeting transcript.
- Don’t describe the meal you want to cook. List the eleven things currently in your fridge.
Long input isn’t a problem. The models handle a full report or a 40-minute transcript comfortably, and the messy specifics inside that material are exactly what makes the reply sound like it understands your situation rather than a generic version of it. Capabilities have moved fast since 2023, which is why the good, the bad and the absurd of ChatGPT’s evolution is worth a read if you wrote the tool off after one disappointing try.
Treat the first answer as a draft, not a verdict
Nobody writes a good email in one pass, and ChatGPT doesn’t either. The people who get strong results treat the first response as raw material and then push back line by line.
Say you’re writing a product description for a standing desk. Turn one: “Rewrite this for someone who’s 45, works from home and has lower back pain.” Turn two: “Better. Cut the word ‘ergonomic’ completely, drop the last sentence, and add one line about setup time.”
Each follow-up takes ten seconds and saves you a manual edit later. If a reply is 80% right, don’t start over. Just say which 20% is wrong.
Ask for options before you ask for paragraphs
Bad first drafts usually mean a bad direction rather than bad writing, so ask for directions first.
“Give me five possible angles for a blog post about meal prep for shift workers. One line each, no explanations.”
Pick two. Then: “Expand angle three into an outline with six headers.” Then write. You’ll spend two minutes choosing and twenty minutes drafting, instead of an hour rewriting something you never liked in the first place. The same move works for subject lines, product names, gift ideas, and naming a database column.
A four-turn example, start to finish
A freelance web designer wants to approach a 12-person architecture practice whose site last changed in 2019. Here’s the exchange, close to verbatim.
Turn 1. “I’m a freelance web designer. I want to email a 12-person architecture practice whose website hasn’t been updated since 2019. Give me three subject lines and a 120-word email. No exclamation marks, no ‘I hope this finds you well’, and mention one specific thing about their portfolio page.”
Turn 2. “Paragraph two is too salesy. Rewrite it so it leads with their problem, not with me.”
Turn 3. “Cut the whole thing to 80 words without losing the specific detail.”
Turn 4. “Now a LinkedIn version, under 50 words, same tone.”
Four turns, ninety seconds, and out comes a short, specific message instead of a template with a company name dropped into the blank. That’s the difference between landing in an inbox and landing in the delete pile.
Where ChatGPT quietly falls apart
It agrees with you too easily
The tool is at its worst when it’s mirroring you. Push a belief at it and it will often reflect it straight back, which is harmless when you’re brainstorming slogans and genuinely dangerous when the subject is medical, legal or financial. That reflex made headlines when a man told ChatGPT he was feeling delusional and was told he was Jesus. Extreme case, obviously. The softer version happens constantly: ask “isn’t this plan solid?” and you’ll get a warm review rather than a hard one.
Verification habits that save you
- Ask for the working, not just the number. “Show the steps” catches arithmetic slips that a confident final figure hides.
- Open every source it cites. Some references look perfectly plausible and don’t exist.
- Test your own premise with “what would make this advice wrong?” before you act on it.
- Date-check anything recent. Model knowledge has a cutoff and it won’t always announce it.
Plan around the fine print
Free and paid tiers are not the same product, and the gap is widening. OpenAI has started showing ads on ChatGPT’s free and Go tiers in India, a reasonable trade for a free tool but worth knowing if you’re running client work through it. In Europe, tougher EU regulation has already reshaped how the service handles personal data and how clearly it has to explain itself.
For genuinely sensitive material, some organisations stopped trusting the public version altogether. The US Department of Defense, for instance, stood up its own isolated version of ChatGPT and Grok rather than routing queries through a consumer app.
The practical rule for the rest of us is blunter: if you wouldn’t email it to a stranger, don’t paste it into a free chat window.
Two prompts worth stealing
Save the prompts that work. Not in your head, in a notes file. Within a month you’ll have a dozen, and they’ll outperform any list you find online because they’re shaped around your actual work.
Two that earn their place:
The blunt editor. “Here’s a draft. Read it as a sceptical [target reader] with no reason to keep reading. List the five weakest sentences in order and explain in one line what’s wrong with each. Don’t rewrite anything yet.”
The Friday review. “Here are my notes from the week: [paste]. Group them into what’s done, what’s stalled and what I’ve been avoiding. For each stalled item, suggest the smallest possible next step.”
Neither is clever. Both give the model a job and real material to work with, and both produce something usable in five minutes. Try the approach on the next thing you’d normally send to a colleague for a second opinion: give it a role, paste the messy original, then push back on the first draft.