Free ai prompt generator for a stronger first draft.

A furniture restorer sells refinished mid century pieces through an online shop and writes every listing by hand. Each one takes twenty minutes and they all end up sounding the same. The task is a usable first draft of a listing description. The inputs: the reader is someone furnishing a first flat, the format is a short paragraph plus three specification lines, and the constraint is that condition must be described honestly because returns are expensive.

The prompt that produces something usable looks like this:

"You are writing a product listing for a small independent shop that restores mid century furniture. Reader: someone furnishing a first flat, interested in the piece rather than the period, and cautious about buying secondhand online. Write a description of the item below in two parts. First, one paragraph of 60 to 80 words covering what it is, what it is made from, what the restoration involved and where it works in a room. Second, three lines headed Dimensions, Condition and Restoration, one sentence each. Describe every mark and repair plainly, including the ones a buyer would only notice in person. Do not use the words stunning, unique, iconic or investment. Do not invent a maker, a date or a provenance that is not in my notes. Match the tone of this line from a listing that sold well: 'A solid teak sideboard with the original brass handles, three shallow scratches on the top that we left alone because sanding them out would have taken the patina with them.' My notes on the item follow below."

That prompt works because it does five things in order: role and context, reader, task with an explicit format, exclusions, then an example of the target tone. Each one closes off a way the output could go wrong. The word range stops it running to three paragraphs. The named forbidden words remove the exact adjectives that would otherwise be deleted by hand from every draft. The instruction against inventing provenance matters most, because a plausible sounding fabricated date in a furniture listing is a real problem for the seller, and it is precisely the kind of gap a model will fill helpfully if you have not closed it.

The weaker version is the one nearly everyone types first: "Write me a product description for a vintage teak sideboard." It returns something readable and useless. With no reader it writes for everyone, so the register lands somewhere between an auction catalogue and a homeware advert. With no length it picks one at random. With no exclusions it opens with stunning and closes with a piece of history. With no notes it will supply a decade and a design influence you never claimed. You then spend four rounds of corrections dragging it back, and because each correction competes with the instructions before it, the final draft is a compromise between briefs rather than the one you meant to give.

Do not write a prompt for work whose answer you cannot judge. Generating something you have no way of checking is not saving time, it is moving the risk somewhere you will not see it. If the output is a technical specification, a legal clause, a medical statement or a set of figures, either you or someone you can ask has to be able to verify it line by line. Where you cannot, ask for the questions to research rather than the finished answer.

Skip the structured prompt when the task is a genuine lookup or a one line transformation. Asking for a capital city, or for a list of dates rewritten in another format, does not need context, audience and constraints wrapped around it. Structure earns its place where the output has a reader and a shape, which mostly means drafting, summarising, rewriting and planning. Over specifying a trivial request just moves the wasted time from the end to the beginning.

Do not reach for a prompt when the thinking has not happened. A brief that says write our positioning statement is asking for a decision that only you hold the inputs for. The same applies to anything requiring information the model cannot have: your prices, your capacity, the reason a client left, what your team is actually good at. Those have to go into the prompt as facts. If you do not know them, the missing piece is research, not phrasing.

Be careful about what you paste in. Client documents, personal data, unreleased plans and anything covered by a confidentiality agreement should not be dropped into a general tool without checking what your organisation permits and what the service does with input. Redact names and figures where the shape of the document is what you actually need. It takes a minute, and it is the difference between a useful habit and one you have to explain later.

A workable order is role and context, reader, task, format, constraints, then examples, with the raw material last. That sequence is not decoration. Context first means every later instruction is read against a settled picture of the situation. Putting the material you want worked on at the end keeps a long prompt readable and makes it obvious where your instructions stop and your data starts, which is the commonest source of confusion in prompts that fail for no visible reason.

Mark that boundary explicitly. A line saying the notes follow below, or a simple delimiter around the pasted text, stops your source material being read as further instruction. It is worth doing every time, and it matters more when the pasted text came from somewhere you do not control, such as a customer email, a scraped page or a transcript, since anything instruction shaped inside it can otherwise be treated as part of the brief. Keep one instruction per sentence. Compound requirements joined by and are the ones most likely to be half followed.

When the output disappoints, work through the causes in order rather than rewriting everything at once. Generic output is almost always a missing audience or a missing point of view. Wrong length is a missing number. Wrong register is a missing example. Confidently wrong facts mean you asked for something outside what you supplied, and the fix is to provide the facts or to ask for uncertainty to be flagged rather than filled in.

Only after those should you conclude that the task itself is a poor fit. Some genuinely are. Anything depending on private information, on very recent events, on precise arithmetic across a long document, or on a judgement that belongs to you, will not improve with better phrasing. Change one element at a time while you diagnose, exactly as you would with an ad test, because rewriting the whole prompt teaches you nothing about which part was carrying the failure. Keep the version that broke next to the version that worked, since the difference between them is the actual lesson.

A search query is a set of keywords aimed at matching documents, so it is short by design and every extra word narrows the results. Prompts behave the other way round: the right added words narrow the output usefully, and stripping a prompt back to keywords removes the very thing that makes an answer specific. People who type prompts like search queries get vague results and conclude the tool is vague.

A system prompt is a different animal again. It sets standing behaviour for a whole session or application: the persona, the boundaries, the house rules, the things that must never appear. It is written once and reused, so it stays general and says nothing about today's task. What you write in the message is the brief for this piece of work: this reader, this material, this format. Mixing them is a common cause of drift. Standing rules buried in a one off message are forgotten a few turns later, and task detail parked in a system prompt keeps leaking into unrelated answers.

The value of a good prompt is in reuse, and most people lose it by leaving it in a chat window they will never scroll back to. Save the ones that produce usable drafts in a plain file with a note of the task they serve, and replace the changing part with a visible placeholder in brackets so it is obvious what to swap. Ten of these, covering the things you write repeatedly, is a real piece of infrastructure.

Maintain them as you would any other template. When a draft needs the same manual correction twice, that correction belongs in the prompt as a constraint rather than in your hands again. Over a few months the exclusion list becomes the sharpest part of the whole thing, because it is a record of your own taste written down. Keep the good outputs too, since your best example of the tone you want is usually something the process already produced. And date the file. Behaviour changes as tools change, and knowing when a prompt was last checked tells you whether a sudden drop in quality is yours or theirs.