How to Write Better ChatGPT Prompts: 15 Techniques That Actually Work
Most prompting advice was written for weaker models and has quietly expired. Three things still genuinely change the output, one is worth more than the rest combined, and the popular tricks mostly are not.
Most prompting advice you will find was written for models that no longer exist.
That is not a criticism of the people who wrote it. Techniques that genuinely transformed output on the models of two or three years ago — elaborate role assignment, rigid template structures, incantations about taking a deep breath — worked because those models needed the scaffolding. Current models mostly do not. They infer far more from a plain request, and a good deal of the received wisdom is now cargo cult: things people still do because they used to work.
What follows is what still changes the output, ordered by how much it matters, and an honest account of what has expired.
The One That Matters Most
Context beats technique, and it is not close.
The reason "write me a blog post about productivity" produces something forgettable is not that the phrasing is wrong. It is that the request contains no information about what would make the result good. Who reads this? What do they already believe? What is the piece arguing? What would make you reject a draft?
The model cannot know any of that, so it produces the average of everything written on productivity — which is the definition of generic. It did not fail. It answered the question asked.
Compare that with: this is for experienced developers who are sceptical of productivity advice, the argument is that most time-management systems fail because they optimise the wrong bottleneck, and I do not want any morning-routine content. No technique, no template, no role assignment — just the situation. The output changes completely, because for the first time there was something to aim at.
If you take one thing: describe the situation, not the task. Almost every "bad output" problem is actually a missing-context problem wearing a disguise.
The Two Others Worth Having
State what would make it wrong. Negative constraints are unusually effective, because they eliminate the specific failure you keep getting. "Do not open by restating my question." "No bullet lists." "Do not hedge every claim." These are more actionable than positive instructions like "be engaging", which the model has no reliable way to operationalise.
This works best applied after you have seen the failure. Ask, notice what is irritating about the result, name it, and rerun.
Show an example rather than describing one. If you want output in a particular voice or shape, one sample is worth several paragraphs of description. Paste something you have written and ask it to match the register. Paste a well-structured section and ask for the same structure. Models are considerably better at imitation than at following abstract style directions, and this is the technique that has held its value best.
Iteration Beats Construction
This is the habit change that helps most people more than any individual technique.
The instinct is to compose the perfect prompt before sending anything — to front-load every constraint and get it right first time. It feels efficient and it is usually slower, because you are guessing at what will go wrong instead of observing it.
The faster loop: send something adequate, read what comes back, and correct. "Shorter." "Less formal." "Cut the introduction." "You have missed the main objection." Three rounds of that takes under a minute and lands closer than a carefully engineered opening request, because each correction responds to something real.
A useful variant when you genuinely do not know what you want: ask it to interview you. Before writing anything, ask me whatever you need to know to do this well. It will ask three or four questions, and answering them produces the context that would have been missing — without you having to work out in advance what was missing.
What Has Quietly Expired
Elaborate role assignment. "Act as a world-class expert marketer with 20 years of experience" did real work on older models. On current ones it mostly does not, because the model already infers the appropriate expertise level from the request. It retains some value for genuinely unusual perspectives — asking for a specific and unexpected viewpoint — and very little for standard professional tasks.
Politeness and pressure tactics. Offering tips, claiming urgency, emotional framing. These circulated widely, produced inconsistent results in testing, and are not worth building habits around.
Rigid templates. The multi-part frameworks with mandatory sections impose structure that often has nothing to do with your actual request. They help beginners by forcing them to supply context, which is real but incidental — the context was the active ingredient, not the template.
Long incantations generally. The trend runs the other way. Each generation needs less scaffolding, because reducing the need for it is exactly what the labs are working on.
That last point has a career implication worth stating plainly: prompting is becoming literacy rather than expertise. Worth learning, not worth building a professional identity around. The deeper treatment is in the prompt engineering guide, and the broader argument about which AI skills hold value is in the future of work.
Set It Once Instead
If you find yourself repeating the same context every conversation — your role, your industry, your style preferences — you are doing manually what custom instructions do automatically.
Configure them once and every subsequent conversation starts with that context already in place. It is the highest-return ten minutes available in the entire tool and most people never open the setting. The custom instructions guide covers what actually belongs in there.
The Short Version
Describe the situation rather than the task. Name the failure you want to avoid. Show an example instead of describing a style. Iterate rather than constructing. Put your standing context in custom instructions so you stop retyping it.
That is essentially all of it. The rest is either obsolete, marginal, or someone selling a framework — and the models keep moving in the direction of needing less of it, not more.
For applying this at work specifically, how to use ChatGPT for work covers where the output is reliable and where it is not, and ChatGPT tips and tricks covers the product features — memory, projects, file handling — that change results more than prompt phrasing now does.
Frequently Asked Questions
- Why does ChatGPT give generic responses?
- Almost always because the request was generic. Asked for "a blog post about productivity" the model has no way to know your audience, your angle, or what you would consider good, so it produces the statistical average of everything it has read on the topic. The average is exactly what generic means.
- What is the best ChatGPT prompt format?
- There is no best format, and the templates promising one are mostly selling structure for its own sake. What consistently matters is context — who the output is for, what situation it sits in, and what would make it wrong. Supply those in plain language and the formatting takes care of itself.
- Do role prompts like "act as an expert" still work?
- Much less than they used to. On older models the role framing meaningfully shifted output quality; current models mostly infer the appropriate register from the request itself. It still helps for genuinely unusual perspectives, and it is largely superstition for ordinary professional tasks.
- Is prompt engineering still a useful skill?
- As a literacy, yes. As a profession, it is eroding by design — every model release reduces how much prompting technique is needed, because inferring intent is precisely what the labs are optimising. Learn it, use it, and do not build a career identity on it.
- What is the fastest way to improve my results?
- Stop trying to write the perfect first prompt and start iterating. Two rounds of "shorter, less formal, cut the introduction" will beat twenty minutes of prompt construction almost every time, because you are correcting against real output instead of guessing in advance.
