Most people use ChatGPT like a search box with better manners. Type a question, read the answer, close the tab. Fine for trivia. It falls apart the moment you need something with nuance, like a client email that has to land a certain way or a plan you’ll actually follow through on.
The gap between a mediocre answer and a genuinely useful one is rarely about the model. It’s the setup, the follow-ups, and a handful of habits that take about ten minutes to learn. Here are the ones worth keeping.
Set your custom instructions once and never think about them again
There’s a settings box most people never open. It has two fields: one about you, one about how you want responses. Filling them in properly removes the need to re-explain your context in every new chat.
A real example that works better than “be concise”:
I’m a marketing manager at a 40-person SaaS company. I write for technical buyers who dislike hype. Default to short paragraphs, UK spelling, and numbers over adjectives. If I ask for copy, give me one strong option rather than five weak ones. Flag any claim you can’t back up.
That last sentence does more work than it looks like. It nudges the model toward honesty instead of confident filler.
Name the role and the reader in the same breath
Vague prompts get vague answers. Compare “write an email about our price increase” with this:
You’re a customer success lead writing to a client who’s been with us for four years about a 7% price rise effective 1 April. They’re price-sensitive and complained once already. Keep it under 150 words, don’t apologise more than once, and don’t offer a discount.
Same request, wildly different output. The extra detail costs you thirty seconds and saves you three rounds of editing.
Treat the first response as a rough draft
Nobody reads the first answer and publishes it. The people getting good results are mostly getting better at revision. “Make it better” changes almost nothing. Specific instructions do.
Follow-ups that reliably move the needle:
- “Cut 30% without losing the argument.” Vague length requests get ignored; a percentage does not.
- “Where is this weakest? Argue against yourself.” Useful for spotting the paragraph you were hoping nobody would read.
- “Give me three versions: blunt, warm, formal.” Faster than describing the tone you want.
- “Replace every adjective with a fact or a number.” Kills the word “seamless” instantly.
- “What would a sceptical reader ask next?” Surfaces the objection you forgot to address.
Let it interview you before it writes
For anything fiddly, a short detour beats a long rewrite. Try: “I need a 100-word speaker bio. Ask me eight questions, one at a time, then write it.” The questions are often better than the ones you’d have thought to answer, and you end up with material you can reuse.
Upload the actual document instead of describing it
Pasting a summary of a contract, a meeting transcript, or a spreadsheet and asking questions about it is one of the most underused capabilities. Grounding the answer in the real file beats asking it to recall general knowledge by a wide margin.
Two habits make it safer. Ask it to quote the exact line it’s relying on, and ask what it couldn’t find. Long documents get skimmed internally, and the model will happily fill gaps with something plausible rather than admit the section was cut off.
Keep a prompt file and a couple of Projects
When a prompt produces something you’d use again, save it. A plain text file with a dozen tested prompts is worth more than any list of tricks you read once and forget. Group them by task: cold outreach, performance review drafts, weekly updates, interview prep.
Projects work well for anything recurring. Dump in your brand guidelines, your last three quarterly reports, and your tone-of-voice doc, and set the instructions once. Every chat inside that project starts with the context already loaded.
Verify every number, name and citation
It will invent a statistic and attach a real-sounding source to it. This is not a rare edge case. Treat any figure, quote, date, or reference as unverified until you’ve checked it yourself.
A quick sanity check: ask “which parts of this answer are you least confident about?” It often flags exactly the sentence you should have doubted. Then check that sentence manually anyway.
Use voice mode for messy thinking
Talking is faster than typing when your thoughts aren’t organised yet. Ramble for ninety seconds about the problem you’re stuck on, then ask for the three main threads and a suggested order. It’s a surprisingly good way to get unstuck before you start writing.
Know when a different tool is the better call
ChatGPT is one option among several, and the differences between models matter for specific jobs. Long document analysis, current-events questions, and code generation don’t all favour the same product, so it’s worth keeping a second option within reach. If you want to compare answers side by side rather than opening six tabs, Poe lets you run several major chatbots from one place, which makes the comparison quick. There’s also a broader roundup of AI chatbots online that are actually worth your time if you’re still choosing.
Keeping your own voice when the draft comes from a model
Generic output is the most common complaint, and it usually comes from generic input. Feed it two or three samples of your own writing and ask it to match the rhythm rather than the vocabulary. Then rewrite the opening line yourself. If you want a longer walkthrough of that workflow, this piece on getting real value from AI text generators without losing your voice covers it in more depth.
For line-level polish on something you’ve already written, a dedicated rewriting tool is often the better fit than a chat window. This Wordtune review explains where that trade-off makes sense.
If you’re letting it write code
Chat-generated scripts are fine until they touch a real database or a live server. Give it read-only access first, keep changes in small commits, and read every line before it runs. The guide to running coding agents safely is worth reading before you hand over anything that can delete a file.
Start with custom instructions and the follow-up list. Those two changes alone will do more for your output this week than a hundred bookmarked prompt collections.