Ask ChatGPT to “write a blog post about coffee” and you’ll get 600 words that are technically correct and instantly forgettable. Ask it to “rewrite this paragraph for a reader who already owns a grinder and doesn’t need the basics explained” and something different happens. The model didn’t get smarter between those two prompts. You just gave it something to aim at.
That’s most of the game. ChatGPT is a text-prediction engine that happens to be unusually good at following instructions, which means the quality of what comes out is largely a function of what you put in, plus how willing you are to argue with the first draft.
What the model is actually doing
It isn’t a database, and treating it like one is where people get burned. It doesn’t retrieve a stored answer and hand it over. It generates the most plausible continuation of your text, token by token, based on patterns in training data. Usually that produces something accurate. Occasionally it produces something fluent, confident, and wrong, which is worse than an obvious error because it reads so well.
The practical consequence: ChatGPT is strongest when the task is about language, structure, or transformation, and weakest when the task requires a precise recall of a fact it may have absorbed imperfectly. Ask it to reorganise your thinking and it shines. Ask it to tell you the exact wording of a 1997 regulation and verify it yourself.
Where it genuinely earns its keep
After a couple of years of daily use, the tasks that reliably save time are narrower and less glamorous than the demos suggest:
- Converting messy notes into a structured draft you can edit into shape
- Rewriting the same idea in four different tones so you can steal the best sentence from each
- Explaining a dense concept at three levels of difficulty in a row
- Producing the scaffolding nobody wants to write: agendas, checklists, table columns, subject lines, error messages
- Playing devil’s advocate on a decision you’ve already half-made and want stress-tested
Notice what’s missing. Original research. Judgement calls. Anything where being wrong costs real money and you can’t easily check the answer.
Three ingredients that separate a useful prompt from a wasted one
Context it can’t guess
Who’s reading, what they already know, what you’ve tried, what failed. “Write a follow-up email to a client” is a coin flip. “Write a follow-up to a client who went quiet nine days after a proposal, who mentioned budget approval in March, and who hates being chased” gives it a target. Two extra sentences of context usually beat another paragraph of instructions.
A format, stated plainly
If you want a table, say table and name the columns. If you want under 150 words, say under 150 words. If you want no bullet points, say that. Left unspecified, you get the average of every format it has ever seen, which is often a wall of paragraphs with a bolded intro line.
A constraint that forces a choice
Constraints do more work than adjectives. “Make it punchy” is meaningless. “No sentence longer than 15 words, no adverbs, no opening with a question” forces decisions. So does picking a side: “Argue the opposite of what I just said and don’t hedge.”
The rough-notes workflow worth stealing
The single most useful pattern I’ve found is treating ChatGPT as a finishing tool rather than a starting one. You do the thinking in your own disjointed, half-sentence way, then hand over the mess. There’s a five-step workflow that turns rough notes into finished work which lays this out in detail, and the core idea is simple: dump everything, ask for structure only, then rewrite the parts that sound like nobody. The model is good at shape. You keep the voice.
When the task has too many moving parts
Logistics is where people either love ChatGPT or give up on it. A birthday dinner for twelve with two vegetarians and a nut allergy is a problem it handles well, provided you feed it the constraints in one go rather than dribbling them out. For a longer demonstration, this eight-step walkthrough builds an entire street party plan from a single messy brief, including the permits, the budget split, and the rain contingency. The takeaway isn’t the party. It’s the sequence: scope, constraints, draft, critique, revise.
Where it still falls apart
Long documents are the weak spot. Give it a 40-page contract and ask for every clause mentioning liability and you’ll get a confident list that quietly omits two. Numbers spanning a table drift. Citations are a genuine hazard: it will invent a plausible-looking source with a real author, a real journal, and a page number that doesn’t exist. If a claim matters, verify it outside the chat window.
The other failure mode is tone collapse. Left alone, output tends to drift toward a smooth, slightly over-friendly register with a lot of tricolons and a summary paragraph nobody asked for. A breakdown of where the chatbot gets it right and where it gets it wrong is worth reading if you keep hitting that same flatness; the fix is nearly always tighter instructions about audience and banned phrases, not a different model.
Push back instead of starting over
Most people’s instinct when a response is mediocre is to delete the whole thing and rephrase from scratch. That throws away useful material. Better: reply with a specific complaint. “Paragraph three is generic. Rewrite it with a concrete example involving a small bakery. Keep the rest.” The model handles surgical corrections well, and you can iterate four or five times in the time it takes to compose one perfect mega-prompt.
Two follow-ups that earn their keep: “What did you assume that I didn’t tell you?” and “What’s the weakest claim in this?” Both surface the places where you’d otherwise publish something shaky.
The argument underneath all of this
There’s a live disagreement about whether this whole approach is the right one. The current chat paradigm depends on you supplying context and the model filling gaps. Mira Murati, who led the team that built ChatGPT, is now betting against the way it works, pushing toward systems that collaborate rather than respond. Whether or not that pans out, the useful lesson for now is that you are not using a finished product. You’re using an interface that will look strange in five years.
Keep a small test set
Here’s a habit that pays off more than any prompt trick. Save three or four real tasks you do regularly, with your actual inputs and what a good output looks like. Mine are a product description rewrite, a meeting summary, and a short factual explainer on a technical topic.
Run them whenever a new model lands. It takes ten minutes, and it tells you something no benchmark or launch post will: whether the thing is genuinely better at your job, or just better at sounding like it is. Models change. Your test set is the only honest measure you’ll have.