How to Write AI Image Prompts That Actually Work

A Kanakaw guide · about 10 min read

The Anatomy of a Prompt That Works

Most weak prompts fail for the same reason: they describe an idea, not an image. "A beautiful woman" is an idea. A prompt that works reads more like a shot list — it tells the generator what is in the frame, where it is, how it is lit, and how it should look. A useful way to build one is to cover five parts in order: subject, action or pose, setting, light, and style.

Start with the subject and be specific. Not "a dog" but "a golden retriever puppy, tongue out, ears flopping". Not "a city" but "a narrow street in Lisbon at dusk, trams, laundry lines overhead". Then add what the subject is doing, then the setting, then the light, then the look. You do not need all five every time, but if your result feels random, the missing part is usually the one you skipped.

Length matters less than order. Prompts on Kanakaw can be up to 3,000 characters, but the sweet spot for most images is roughly 40 to 120 words. Shorter than that and the generator fills in the gaps with its own guesses; much longer and the extra words start fighting each other. Put the most important details first — the subject and the composition — and the finishing touches last. If two ideas conflict, the one written first usually wins.

One more habit that helps: change one thing at a time. Change the light, keep everything else. Change the lens, keep everything else. Changing five things at once teaches you nothing about which one fixed the picture.

How to Describe Light, Camera and Lens

Light is the single most powerful word in a prompt. It decides the mood, the colours and how realistic the result feels. "Golden hour" gives you low, warm sunlight and long shadows; "soft window light" gives you gentle, even illumination on a face; "overcast daylight" gives you flat, honest light with no hard shadows; "neon at night" gives you colour and contrast. Name the light the way a photographer would, and name where it comes from: "morning light through a large window, from the left" beats "nice lighting" every time.

Camera language works the same way. A few terms carry a lot of meaning. "85mm portrait lens" means a flattering, slightly compressed face with a blurred background. "24mm wide-angle" means an expansive landscape or a room that feels big. "Macro" means extreme close-up with fine detail. "Aerial view" or "drone shot" means looking down from above. "Shallow depth of field" blurs the background and pulls the eye to the subject; "everything in focus" does the opposite and suits architecture and landscapes.

Shot type matters too. "Close-up", "medium shot" and "full body" decide how much of the subject is in the frame, and "eye level", "low angle" and "bird's-eye view" decide where the camera stands. A portrait described as "low angle, looking up" feels powerful; the same person described as "high angle, looking down" feels small and quiet. If the framing keeps coming out wrong, the fix is usually one of these words, not a longer prompt.

Finally, match the shape to the shot. A sweeping landscape wants 16:9, a phone wallpaper or social portrait wants 9:16, a poster wants 2:3, and a profile picture wants 1:1. Kanakaw offers 16:9, 9:16, 1:1, 4:3, 3:4 and 2:3 — pick the ratio before you write the prompt, because a scene composed for a wide frame will not survive being squeezed into a square.

Style Keywords That Actually Change the Picture

Style words are the fastest way to change the whole feel of an image, but only if you use the right ones. For photography, the words that work are the ones photographers use: "editorial", "candid", "documentary", "studio portrait", "film grain", "35mm film look", "black and white", "high-key" for bright and airy, "low-key" for dark and moody. For illustration, name the medium: "watercolour", "oil painting", "pencil sketch", "flat vector", "pixel art", "anime", "storybook illustration". The generator knows what these look like; vague words like "beautiful", "amazing" and "high quality" do not add anything, because every prompt already asks for a good picture.

Two or three style words are enough. Piling on ten — "cinematic, hyper-realistic, ultra-detailed, 8k, award-winning, masterpiece" — usually makes the image busier and less coherent, not better. Pick the one style that defines the look, add one texture word if it helps ("matte", "glossy", "grainy", "soft focus"), and stop.

Colour words deserve the same discipline. Instead of "colourful", name the palette: "warm earth tones", "muted sage green and cream", "cold blue and steel grey". A named palette keeps the whole image consistent, while "colourful" lets every part of the image pick its own.

One practical note: Kanakaw chooses the image engine behind the scenes, so you never need to write engine names into a prompt. You only choose between two tiers in the generator — Standard at 6 credits per image, and Premium at 41 credits, which is members-only. The words in your prompt do the same job whichever tier you use; Premium simply renders them with more detail.

Negative Prompts and Fixing Common Failures

A negative prompt tells the generator what to leave out. Kanakaw already starts you off with a sensible default — "bad quality, blurry, watermark, text" — which handles the most common junk. Add to it only when you have a specific problem: "extra fingers, deformed hands" for hands, "distorted face, duplicate, cropped" for portraits, "oversaturated, harsh shadows" for colour problems. A negative prompt is a scalpel, not a second prompt: if you write a paragraph of negatives, you start fighting the positive prompt too.

Hands are the classic failure: they are complicated, and a vague prompt leaves the generator guessing. Three fixes help, in this order. First, describe the hands in the prompt itself — "hands resting in her lap" or "holding the mug with both hands" — because a described hand renders better than an undescribed one. Second, add "deformed hands, extra fingers" to the negative prompt. Third, simplify: if the hands still fail, reframe the shot so the hands are out of frame, for example a close-up portrait instead of a full-body shot.

Garbled text is the second classic failure. The rule is simple: do not ask the generator to render words unless you have to. If a sign or a label must be readable, keep it to one or two short words, describe it plainly ("a small chalkboard sign reading OPEN"), and let the default negative prompt's "text" entry suppress the rest. If the text still comes out wrong, the reliable fix is to generate the image without any text and add the words afterwards in any basic editor.

Muddy colours usually mean the prompt asked for too many things at once, or for colours that fight each other. Cut the prompt back to one palette and one light source, and the colours clean up. If one part of an otherwise good image is wrong — a stray object, a warped corner, a bad hand — do not regenerate the whole thing: inpainting (21 credits, members-only) lets you repaint just that spot, and the image upscaler (61 credits, members-only) sharpens a good image up to four times larger.

And iterate cheaply. Draft in Standard at 6 credits a time, asking for up to 4 variations of one prompt so you can compare compositions side by side. Once the composition, light and framing are right, run the same prompt once in Premium at 41 credits for the finished version. Cheap drafts, one expensive finish — that is how you explore freely without burning through your balance.

Before and After: Four Prompts Rewritten

Here are four weak prompts, why each one fails, and the stronger version that fixes it.

1. Portrait. Weak: "a beautiful woman". It says nothing about age, pose, clothes, place or light, so the generator guesses all of it — and the guesses are usually generic. Strong: "Editorial portrait of a woman in her late twenties, shoulder-length dark hair, soft smile, looking slightly off camera, cream linen shirt, sitting by a large window in a bright apartment, soft morning light across her face, 85mm portrait lens, shallow depth of field, warm muted tones, candid, photorealistic". Every addition answers a question the weak prompt left open.

2. Product. Weak: "a coffee mug". A floating object on a blank background, because no setting was given. Strong: "Product photograph of a matte ceramic coffee mug in sage green, steam rising, on a light oak table next to a small vase of dried grasses, soft window light from the left, minimal Scandinavian styling, 50mm lens, crisp focus on the mug, warm neutral background, advertising photography". The setting and the light turn an object into a photograph.

3. Landscape. Weak: "a mountain". One word of subject and nothing else, so the result is a postcard cliché. Strong: "Alpine lake at dawn, still water reflecting snow-capped peaks, a wooden jetty leading into the frame, low mist over the water, first light turning the clouds pink and gold, wide-angle 24mm, rich detail, calm and quiet mood, landscape photography". The jetty gives the composition a leading line; the time of day gives it colour.

4. Pet. Weak: "a cute dog". "Cute" is not a visual instruction, and the generator has no idea what the dog is doing or where it is. Strong: "Golden retriever puppy sitting in tall autumn grass, golden leaves falling around it, tongue out, soft late-afternoon backlight, 85mm lens, shallow depth of field, warm golden tones, joyful and candid, pet photography". Breed, action, setting, light and lens — five answers, one picture.

Notice the pattern across all four: the strong prompt never uses more adjectives, it uses more facts. Subject, action, setting, light, lens, palette — write those six things and the generator has almost nothing left to guess.

Frequently Asked Questions

How long should an AI image prompt be?

Most good prompts are about 40 to 120 words. That is enough to cover the subject, the setting, the light and the style without the details starting to fight each other. Prompts on Kanakaw can be up to 3,000 characters, but longer is not better — if a prompt is not working, cutting it back to the essential facts usually helps more than adding more words.

Why do my images keep coming out with wrong hands?

Hands are one of the hardest things for image generators, and vague prompts make it worse. Describe the hands in the prompt itself, such as "hands resting in her lap", and add "deformed hands, extra fingers" to the negative prompt. If the hands still fail, reframe the shot so they are out of frame, or fix just that spot afterwards with inpainting, which costs 21 credits and is members-only.

What should I put in a negative prompt?

Start with the default, "bad quality, blurry, watermark, text", which Kanakaw already fills in for you. Add only the specific flaw you are trying to avoid, such as "extra fingers, deformed hands" or "oversaturated, harsh shadows". A short, targeted negative prompt works better than a long list, which can start fighting the main prompt.

Should I draft in Standard or go straight to Premium?

Draft in Standard at 6 credits per image. Generate up to 4 variations of one prompt and compare them side by side to find the composition and framing you like. Once the draft is right, run the same prompt once in Premium at 41 credits for the finished version. Premium is members-only, and membership comes with any credit pack — 30 days with a monthly pack, 12 months with an annual pack.

Can I use my generated images commercially?

Yes. Subject to the Terms of Service, you may use the images you generate for personal and commercial projects. The same rule covers images you restyle from your own photos with image-to-image (49 credits, members-only) and images you fix up with inpainting or the upscaler.

Try it: AI Image Generator · Pricing · All guides