You type “a beautiful sunset over mountains” and get something that looks like a stock photo from 2009. So you add “amazing” and “high quality” and “8k” and “masterpiece”, and it gets marginally shinier and no better.

The problem is not that you need more adjectives. It is that adjectives like “beautiful” carry almost no information. The model cannot act on them, so it falls back on the most average interpretation of your subject that exists in its training data.

Good prompting is not about magic words. It is about replacing vague quality claims with specific decisions.

Specificity beats intensity

This is the single rule that fixes most prompts. Every time you are tempted to add an intensifier, add a detail instead.

“A beautiful sunset over mountains” gives the model nothing to work with. “Low sun behind jagged granite peaks, orange light raking across a snowfield in the foreground, thin cloud catching pink above the ridge” describes an actual image. The second prompt is not more enthusiastic, it is more decided.

The test is simple: could two people read your prompt and picture noticeably different images? If yes, the model is choosing for you, and it will choose the average.

Words that do nothing

A set of terms circulate in prompt guides that mostly serve as reassurance rather than instruction.

• “Masterpiece”, “award-winning”, “best quality”. These describe your hopes, not the image. Modern models already aim for coherent output.
• “8k”, “ultra HD”, “highly detailed”. Resolution is set by your generation settings, not by asking. These sometimes nudge toward a glossy digital-art look, which may not be what you wanted.
• “Trending on ArtStation”. A relic of earlier models. It mostly pulls toward a particular era of digital fantasy illustration.
• Long strings of comma-separated adjectives. Attention gets diluted. Twenty modifiers usually produces a muddier result than five specific ones.

None of these are harmful exactly. They just occupy space that a real decision could occupy.

Style is a decision, not a compliment

“Artistic” means nothing. “Charcoal drawing on textured paper, heavy smudging, high contrast” means something the model can execute.

Naming a medium is the highest-leverage single change you can make to a prompt. Oil painting, watercolour, 35mm film photograph, vector illustration, clay render, pencil sketch — each carries an entire visual grammar with it, and each will transform the same subject completely.

A note on naming living artists: it raises real ethical and legal questions, and some platforms restrict it. Describing the visual qualities you want — the brushwork, the palette, the era — gets you most of the way there without appropriating a specific working artist’s name.

Composition and light do the heavy lifting

Most people describe what is in the image and never describe how it is seen. That is why generated images so often feel flat and centred — because nothing told the model otherwise.

• Distance. Extreme close-up, medium shot, wide establishing shot. Changes the emotional register entirely.
• Angle. Low angle looking up, overhead, eye level. A low angle makes a subject imposing without you saying “imposing”.
• Light source and quality. Soft window light from the left, harsh midday sun, single candle, neon signage reflecting off wet tarmac.
• Time of day. Golden hour, blue hour, overcast noon. Each brings its own color temperature and shadow behavior.

If you change nothing else about how you prompt, add a light description. It is consistently the difference between an image that looks generated and one that looks photographed or painted.

Negative prompting, where supported

Some tools accept a separate field for what you do not want. Where available it is useful, and it works best for concrete objects rather than abstract qualities.

“No text, no watermark, no extra limbs” is actionable. “Not ugly, not boring” is not, because the model has no stable representation of either.

Where there is no negative field, phrase things positively in the main prompt. “Empty street” works better than “street with no people”, because mentioning people at all raises the chance they appear.

Iterate on one variable

The most common workflow mistake is rewriting the whole prompt when the result disappoints. You then have no idea which change helped.

1. Start deliberately simple — subject, action, context. Generate.
2. Add the style. Generate again. Keep it if it moved toward what you wanted.
3. Add lighting and composition. Generate.
4. Only then adjust technical detail.

Keep the prompts that worked. A personal library of proven starting points is worth far more than any list of magic keywords, because it is calibrated to the tools you actually use and the results you actually like.

If you would rather start from prompts that are already tested, our Android app GenZ AI Prompts is a curated library organized by category, built for exactly this — copy something proven and adjust from there rather than starting from an empty box.

Worked example: rebuilding a weak prompt

It is easier to see this as a sequence than as a rule. Start with the kind of prompt most people write:

“A beautiful woman, amazing quality, 8k, masterpiece, highly detailed”

There is one piece of information here — a woman — and four expressions of hope. The model has to invent everything else, so it produces the statistical average of every portrait it has seen.

Add subject specifics. “A woman in her sixties with short grey hair and deep laugh lines.” Now there is a person rather than a placeholder.

Add action and context. “…sitting at a cluttered kitchen table, hands wrapped around a mug of tea, mid-conversation.” The image now has a situation, which is what makes pictures feel real.

Add style. “…shot on 35mm film, slight grain.” This one word — film — does more than every quality adjective in the original combined.

Add light and framing. “…soft late-afternoon light through a window to her left, medium close-up at eye level, shallow depth of field.”

The finished prompt contains no quality claims at all, and it will outperform the original substantially. Every word has been replaced with a decision.

Where models still struggle

Knowing the failure modes saves you from blaming your prompt for something the tool cannot currently do.

• Text inside images. Improving fast, but still unreliable for anything beyond a few words. If accurate text matters, add it afterwards in an editor.
• Counting. “Exactly five birds” is a request, not a guarantee. Models handle small numbers better than large ones and often ignore the count entirely.
• Spatial relationships. “The red cup to the left of the blue book” frequently comes out swapped. Simplify the scene or generate elements separately.
• Consistent characters across images. Getting the same face twice is genuinely hard without dedicated features built for it.
• Hands and complex anatomy. Much improved, still the first place to check before you call an image finished.

When a prompt keeps failing on one of these, the answer is usually to change the composition so the difficulty disappears rather than to keep rewording. Framing a portrait so the hands are out of shot solves the hands problem completely.

Prompts for text models are a different craft

Everything above concerns image generation. Writing prompts for a language model follows different rules, and conflating the two produces poor results in both directions.

• Give context, not just an instruction. Who the output is for and what it needs to achieve changes the answer far more than adjectives do.
• Show an example. One sample of the output you want communicates more than three paragraphs describing it.
• State the constraints. Length, format, tone, what to leave out. Unstated constraints get filled with defaults you did not choose.
• Ask for reasoning on hard problems. Requesting the working before the answer measurably improves accuracy on anything multi-step.
• Iterate conversationally. The first output is a draft. Refining it in dialogue beats trying to write one perfect instruction.

Frequently asked questions

How long should a prompt be?

Long enough to remove ambiguity, short enough that every word is doing work. For images, twenty to sixty words of specific description usually outperforms both a five-word prompt and a two-hundred-word one.

Do prompts transfer between tools?

Descriptive language transfers well. Tool-specific syntax and parameters do not. A prompt written in plain description will work reasonably across generators; one built around one platform’s flags will not.

Why do I get a different image each time?

Generation is random by design, seeded differently each run. Some tools let you fix the seed, which gives you a reproducible starting point so you can change one word and see only that word’s effect.

Is prompt engineering going to stay useful?

The specific tricks age quickly — half the advice from two years ago is now useless. What does not age is knowing what you want and being able to describe it precisely. That skill predates AI and will outlast any particular model.

The underlying skill

Prompting well is mostly art direction. Photographers and illustrators have always had to decide the subject, the framing, the light and the treatment before they could produce anything. AI has not removed that requirement, it has just moved it into a text box.

Which means the people who get consistently good results are not the ones who memorized keyword lists. They are the ones who looked at enough images to know what they actually wanted, and then said so.