How to Write AI Image Prompts That Actually Work

TL;DR: What You Will Learn

• Every strong prompt has an anatomy: subject, style, composition, lighting, and technical details.
• Negative prompts tell the AI what to leave out, which fixes most warped and cluttered results.
• Good images come from iteration, not luck. You refine in small, deliberate steps.
• Brand consistency is achievable with reusable prompt templates, reference styles, and locked-in visual rules.

This is written for in-house marketers, solopreneurs, and small business owners in Singapore who are new to AI image tools and just want results they can use.

Why Most AI Image Prompts Fail Before You Even Hit Generate

Let me start with a scene I have watched play out more times than I can count.

Imagine you run a small hawker-style F&B startup selling nasi lemak sets online. You want mouth-watering shots for your Shopee listing and your Instagram store, so you open an AI image tool and type “nasi lemak, nice photo”.

You hit generate, and what comes back looks like a generic Western stock plate: fluffy rice that looks nothing like coconut rice, a fried egg that could be from anywhere, and a plating style you would never see at a real coffee shop in Singapore.

The image is not broken. Your instructions were.

The Model Fills Every Gap You Leave

Here is the thing most beginners miss when learning how to write AI image prompts: the model does not know what you meant, only what you typed.

When you leave the subject vague, the style unspecified, and the context blank, the AI does not pause to ask you. It simply guesses, and its guesses are pulled from an average of everything it has seen.

So “nice photo” becomes generic studio lighting. “Nasi lemak” becomes whatever the model thinks rice and egg look like, usually leaning bland and international rather than local and specific.

The Four Usual Culprits

Across the prompts I have reviewed and rewritten, the same four failures come up again and again.

• Vague subjects: You say “food” when you mean a banana leaf plate of nasi lemak with crispy ikan bilis, sambal, and a fried egg.
• No style direction: You never told it whether you wanted a bright flat-lay for Instagram or a moody close-up for a Shopee banner.
• Conflicting instructions: You ask for “minimalist” and “packed with garnish” in the same breath, so the model splits the difference and satisfies neither.
• Unrealistic expectations: You expect it to nail a specific hawker stall you have in mind without describing it at all.

From a Lazy Line to a Considered Brief

Compare “nasi lemak, nice photo” with something like “top-down flat-lay of a nasi lemak set on a banana leaf, fragrant coconut rice, crispy ikan bilis, red sambal, sunny-side-up egg, cucumber slices, natural daylight, bright and appetising, styled for an Instagram food post”.

The second one leaves the model far less to invent, so you get something usable instead of something generic.

This is the real skill: writing clear, specific instructions a machine can follow. It is the exact discipline behind prompt engineering best practices every beginner should know, and the good news is that it is entirely learnable.

The Anatomy of an AI Image Prompt That Actually Works

When I first started generating images, I typed things like “bubble tea photo” and wondered why the results looked like stock photos from 2012.

The fix was learning that a strong prompt is built from distinct layers, each doing a specific job. Once you understand those building blocks, you stop guessing and start directing.

To make this concrete, I will use a running example throughout: imagine you run marketing for a homegrown Singapore bubble tea brand, and you need a hero image for your new website.

Subject and Action

The subject is what actually sits in the frame, and the action is what it is doing.

Vague subjects give you vague images. “A drink” could be anything, so instead I would write “a tall clear cup of brown sugar bubble tea with visible tapioca pearls, condensation on the glass, a paper straw resting against the rim”.

Adding action brings it to life: “a hand lifting the cup as pearls swirl inside”. Now the AI has something specific to build around.

Style and Medium

Style tells the model what kind of image you want, and this single choice changes everything.

The same bubble tea subject can be rendered as a glossy commercial photograph, a soft 3D render, a flat vector illustration for a promo card, or a loose watercolour for a more artisanal feel.

For our hero image, I would go with “high-end commercial food photography, shot on a full-frame camera”. If the brand later wanted a playful Instagram sticker set, I would swap in “flat illustration, bold outlines” instead.

Composition and Camera

Composition controls how the shot is framed, and it is where a lot of amateur prompts fall apart.

Tell the model the angle and framing directly: “close-up product shot at a slight three-quarter angle, cup positioned off centre following the rule of thirds, shallow depth of field with a blurred cafe background”.

I also specify the aspect ratio to match the placement, so a website hero gets a wide 16:9 while a story ad gets a vertical 9:16. Getting this right saves you from awkward cropping later.

Lighting and Mood

Lighting is the layer most beginners skip, yet it does the most emotional heavy lifting.

“Soft natural daylight from a window on the left, warm and inviting” reads completely differently from “moody neon glow, pink and teal reflections” or “golden hour backlight with a gentle lens flare”.

For a fresh, everyday bubble tea brand, I would pick the soft daylight option. For a late-night dessert bar concept, the neon would suit better.

Stack all four layers together, and the output transforms from generic to branded. If you want to go deeper, these same principles echo the prompt engineering techniques that improve your AI results across text tools too.

Here is a simple template you can reuse:

[Subject and action], [style and medium], [composition and camera], [lighting and mood], [aspect ratio].

Choosing the Right Style, Medium, and Mood for Your Brand

Here is where I see the biggest gap between beginners and people who actually get usable images. The subject is only half the prompt. The aesthetic layer is where on-brand-ness is won or lost, and it is the part most people leave vague.

Let me show you with two Singapore businesses I could easily imagine briefing. A local fintech startup wants clean, trustworthy visuals for its investor deck. A youth streetwear label wants bold, editorial energy for its Instagram grid. Same instruction, “show a person using our product,” but the style words send them in opposite directions.

Naming a visual style precisely

Vague words like “modern” or “nice” tell the model almost nothing. Precise style references do the heavy lifting.

Pull from three vocabularies:

• Art movements and illustration styles: Bauhaus, flat vector illustration, isometric, watercolour, Japanese ukiyo-e.
• Render engines and 3D looks: Octane render, Unreal Engine, clay render, low-poly.
• Photography and film stocks: Shot on Kodak Portra 400, 35mm, soft studio lighting, golden hour, shallow depth of field.

For fintech, I would reach for “clean editorial photography, soft studio lighting, shot on 35mm, minimal background.” For the streetwear label, “high-contrast flash photography, film grain, editorial magazine styling, dynamic angle.” The same person holding the same phone becomes two entirely different brands.

Matching Mood to Message

Mood is emotional direction, and it should follow your message, not your mood on the day.

If you want to signal trustworthiness, lean into calm, ordered, balanced compositions with even lighting and lots of breathing space. That is why banks and insurers look the way they do.

If you want playful, ask for warm colours, casual poses, and slightly imperfect framing. If you want a premium, ask for restraint: dark backgrounds, single hero objects, dramatic shadows, negative space.

Fintech needs to feel trustworthy and premium. The streetwear label needs to be playful and rebellious. Naming the mood word in your prompt keeps the model from defaulting to a generic stock-photo feel.

Colour and Palette Control

Colour is the fastest way to make an image feel like yours. Do not leave it to chance.

Reference your actual brand colours in the prompt, either by name or hex code. Something like “muted navy and warm cream palette, single orange accent” gives you far more consistency than hoping the model guesses right.

My strongest tip: keep a personal style vocabulary cheat sheet. Every time a prompt produces the look you wanted, save the exact style, mood, and colour words in a note.

Over a few weeks, you build a private reference of phrases that reliably match your brand, so writing a new prompt becomes a matter of assembling proven ingredients rather than starting from a blank page each time.

Using Negative Prompts to Remove What You Don't Want

Most people spend all their energy telling the model what to draw. The quiet secret is that telling it what to leave out often fixes more problems in a single line.

A negative prompt is a list of things you explicitly do not want in the image. It is the underused lever that clears up recurring artefacts like mangled hands, random gibberish text, or a background so busy it steals attention from your subject.

Which Tools Actually Support Negatives

Not every tool handles this the same way, so knowing your platform saves you a lot of frustration.

• Stable Diffusion interfaces like AUTOMATIC1111 and DreamStudio expose a dedicated negative prompt field, and this is where the feature is at its most powerful.
• Midjourney uses the `–no` parameter, so you would add something like `–no text, watermark` to the end of your prompt.
• DALL-E has no separate field, so you describe exclusions inside your main prompt with phrases like “no text or signage, clean uncluttered background.”

If you want to go deeper on why placement and phrasing matter this much, this ties into the wider idea of context engineering versus prompt engineering, where structuring your instructions well matters more than piling on words.

A Property Agency Scenario

Picture a small Singapore property agency preparing listing visuals for a condo unit. They want clean, aspirational interior shots, but their first drafts keep coming back with warped sofa legs, a coffee table that melts into the floor, and shop signage in the background reading nonsense letters.

This is exactly what negatives are for. Adding exclusions clears the warped furniture and the gibberish signage without them having to rewrite the whole scene description.

Here is a copyable starter list you can adapt for interior or product shots:

• Extra fingers, deformed hands, distorted limbs
• Text, watermark, signature, logo
• Gibberish signage, misspelled words
• Cluttered background, clutter, messy
• Warped furniture, distorted objects, melting shapes
• Low quality, blurry, JPEG artefacts

Do Not Over-Stuff Your Negatives

Here is the reassuring part: you rarely need a giant wall of exclusions. Start with three or four negatives that match the actual problems you are seeing, then add more only if the artefacts persist.

Over-stuffing backfires. If you list “shadows, reflections, texture, detail” without thinking, you can strip the very qualities that make an interior shot feel real and warm.

My rule is simple. Fix the specific flaw in front of you, generate again, and adjust. Negative prompts work best as a scalpel, not a sledgehammer.

Iterating Efficiently Instead of Rerolling Blindly

Picture an in-house marketer at a Singapore SME, maybe a homegrown skincare brand, staring down a campaign that goes live Friday. She has a limited pool of paid generation credits and needs three usable hero visuals fast.

The instinct is to smash the generate button and hope. That is rerolling blindly, and it burns credits without teaching you anything about how to write AI image prompts that land.

There is a better way, and it is closer to the test-and-learn discipline good marketers already use for A/B testing than to gambling.

This is not just a personal habit. Ethan Mollick, a Wharton professor whose One Useful Thing newsletter is one of the most widely followed resources on practical AI use, argues that the goal is not to craft “the perfect prompt” but to work with the tool interactively: generate something, then adjust it. In his own image experiments, he describes settling on a Midjourney prompt and then generating again and again, modifying the details each time until the shot felt right. That is exactly the deliberate, one-change-at-a-time loop I am walking you through in this section.

Change One Variable at a Time

When your image is nearly right but not quite, resist the urge to rewrite the whole prompt.

Instead, adjust one thing: the lighting, the camera angle, the background, or the subject’s expression. If you change five things at once and the result improves, you have no idea which change did the work.

Our skincare marketer might keep her subject and composition fixed, then test only “soft morning window light” against “bright studio lighting” for the same bottle shot. Now she knows exactly what shifted, and she can reuse that knowledge next time.

Use Variations, Seeds, and Reference Images

Most tools give you shortcuts that stretch every credit further.

• Variations: When one output is close, generate variations of that specific image rather than starting from scratch.
• Seeds: Lock the seed number so the starting composition stays stable while you tweak wording. In Stable Diffusion, a fixed seed makes generation deterministic, while Midjourney’s own guidance notes that seeds are less reliable across sessions and recommends style references for consistency.
• Reference images: Feed in a brand mood board or a previously approved visual to anchor style, so you are steering rather than rolling dice.

For our deadline scenario, seeds are the quiet hero. Lock the seed on a promising bottle shot, adjust only the wording, and you get three cohesive options that look like a set instead of three unrelated pictures.

Keeping a Prompt Log

This is the habit that separates people who guess from people who improve.

Keep a simple log, even just a spreadsheet, with three columns: the prompt you used, what you changed, and whether the result got better or worse. Over a few campaigns, patterns emerge, and you stop repeating expensive mistakes.

A prompt log turns random luck into repeatable skill, and that is what actually saves credits. You start predicting what a change will do before you spend anything.

This evidence-based mindset is not just about images. It is the same logging and iteration thinking you would apply when you build an AI workflow automation step by step, where documenting what changed is exactly what makes the whole system reliable.

Keeping Every Image On-Brand and Consistent

The single image is easy. The hard part, and the reason most teams give up on AI visuals, is making twenty images across two months look like they came from the same brand.

Picture a Singapore homegrown fashion label pushing out weekly Instagram carousels and EDM banners. Three different people touch the visuals: a marketer, a freelance designer, and an intern covering leave. Without shared rules, you get three different looks, and the feed starts to feel like a jumble sale.

Locking a House Style Prompt Block

The fix I always start with is a house-style prompt block: a fixed chunk of text that every contributor pastes into every generation, then adds the scene on top.

For that fashion label, it might read: “editorial fashion photography, soft natural window light, muted warm tones, shot on 50mm, minimalist studio, calm and premium mood.” The person only swaps in the product and pose.

Lock the constant and vary only the variable, and consistency stops depending on individual taste. This is the same discipline you use when you build a content marketing strategy that boosts your SEO: repeatable systems beat one-off inspiration.

Seeds, Reference Images, and Style References

Prompt text alone will not hold a look together. You need the technical anchors your tool provides.

Most platforms let you reuse a seed number so the model starts from the same random point, which helps keep framing steady within a session. Midjourney notes that seeds are not meant to lock a style across prompts and recommends style references for that (Midjourney documentation). Midjourney has style references and, depending on the version, character or omni references; other tools accept a reference image you upload to steer colour and composition (Midjourney documentation).

• Seeds for repeatable structure within one shoot.
• Style references (an image or code) to carry mood across many prompts.
• Reference images to match an existing product photo or set.

For the retail brand, I would fix one style reference for the whole quarter, so the intern in September lands close to what the marketer made in July.

Colour, Logo, and Typography Guardrails

AI is unreliable with exact hex codes, logos, and text, so treat these as guardrails, not outputs.

Generate the imagery, then drop your real logo and headline in Canva or your design tool afterwards. Never trust the model to render your brand name correctly, and never let it fake a logo.

One PDPA point people forget: If you generate people who resemble real staff or customers, or upload someone’s photo as a reference, you are handling likeness and possibly personal data. Use clearly synthetic faces, or get consent before feeding real photos in.

Finally, save all of this in a shared prompt library, a simple Google Doc or Notion page, so when the freelancer leaves, your house style, seeds, and guardrails stay with the team rather than walking out the door.

Copy and Paste Prompt Templates for Common Marketing Needs

Here is the part you have been waiting for. Once you understand the anatomy of a good prompt, you can stop staring at a blank box and start from a proven structure instead.

Every template below follows the same five slots: subject, style, composition, lighting, and negatives. Fill in the brackets, keep the structure, and you will get usable results far faster than typing a vague one-liner.

The Five Workhorse Templates

These are starting points, not final answers. Run one, read the result, then adjust a single slot at a time.

• Hero product shot: [product] on [surface or background], [style, e.g. clean studio photography], centred composition with generous negative space, [lighting, e.g. soft diffused light from the left], negatives: no clutter, no text, no distorted logos.
• Social media banner: [key message or theme] as a horizontal banner, [style, e.g. flat vector illustration], [subject] positioned to the right leaving room for headline on the left, bright even lighting, negatives: no watermark, no cropped faces.
• Blog featured image: [topic] visualised conceptually, [style, e.g. modern editorial photo], wide 16:9 composition with a clear focal point, [lighting, e.g. natural window light], negatives: no gibberish text, no overcrowding.
• Event poster: [event name and vibe] as a vertical poster, [style, e.g. bold retro graphic], [subject] centred with space at the top and bottom for titles, dramatic [lighting], negatives: no misspelt words, no low resolution.
• Lifestyle scene: [people or activity] in [setting], [style, e.g. candid documentary photography], rule of thirds composition, [lighting, e.g. warm golden hour], negatives: no stiff poses, no extra limbs.

A Worked Example: A Café National Day Promotion

Say you run a small café in Tiong Bahru and you want a hero image for a National Day set menu post.

Start with the hero product template and fill it in: A kaya toast set and iced kopi on a marble table, clean studio photography with subtle red and white bunting in the soft background, centred composition with negative space for a headline, soft diffused morning light, negatives: no clutter, no visible brand logos, no text on the image.

Run it, and you might find the bunting looks too busy. Change one slot: swap the background line for a plain cream backdrop. Run it again.

That single variable swap is the whole game. Keep the structure, change the subject to your croissant, change the style to warm film photography, and you have a matching series without rebuilding the prompt each time.

If you want to speed this up further, keep your filled templates in a document and reuse them, the same way we suggest saving reusable prompts in a ChatGPT user guide for the Singapore workforce.

Treat every generation as a draft. The template gets you to a serious first attempt, and iteration gets you to the image you actually wanted.

Avoiding Legal, Ethical, and Quality Pitfalls in Singapore

Knowing how to write AI image prompts is only half the job. The other half is making sure the image you generate does not land you in legal or reputational trouble, especially if you plan to publish or advertise it.

Let me walk you through the checks I make before any AI image goes live.

Commercial Rights and Tool Licensing

Not every AI tool grants you commercial usage rights by default. Some free tiers only permit personal use, and others reserve rights depending on your subscription plan.

Always read the licensing terms of the specific tool you are using before you monetise an image. Midjourney, Adobe Firefly, and DALL·E each have different terms: OpenAI assigns you ownership of DALL·E outputs and permits commercial use, while Midjourney ties commercial rights to your subscription tier. Firefly in particular markets itself as trained on licensed and public domain content, which many Singapore brands prefer for peace of mind.

Likeness, Trademarks, and PDPA

Here is a scenario I see often. A Singapore agency is about to run a paid Meta ad for a local F&B client, and the creative director loves an AI-generated image of a smiling customer.

Before publishing, they should ask: does this face resemble a real, identifiable person? Does the background contain a recognisable brand logo or trademark that was not cleared?

Generating an image that mimics a real celebrity or public figure can expose you to likeness and publicity claims. Trademarks sneaking into a shopfront or product label can trigger infringement issues too.

The Personal Data Protection Act (PDPA) matters when you edit or generate images of identifiable individuals. If you upload a real customer’s photo into a tool to alter it, you are processing personal data, and you need proper consent and a lawful basis to do so.

Singapore is actively pushing responsible AI adoption, and IMDA’s Model AI Governance Framework, including its generative AI extension, sets out expectations around transparency, accountability, and safe deployment. Treating these principles as a baseline protects both your client and your own credibility, which ties directly into how Google evaluates trust with E-E-A-T.

A Quick Pre-Publish Quality Check

Even a legally clean image can look amateurish. AI models still produce mangled hands, warped text, and inconsistent lighting that undermine your brand.

Before anything goes live, I run through a short checklist:

• Resolution and format: Is the image sharp enough for its placement, whether that is a web banner or a printed EDM?
• Artefacts: Are there distorted hands, gibberish text, extra fingers, or melted objects?
• Human review: Has a real person eyeballed it for tone, accuracy, and cultural fit for a Singapore audience?

That final human review is non-negotiable. AI gets you most of the way, but a trained eye is what keeps your work safe, professional, and genuinely on brand.

Conclusion

If you take one thing away from this guide, let it be this: learning how to write AI image prompts is a structured skill, not a matter of luck or endless lucky rerolls.

We started with the anatomy of a strong prompt: subject, style, composition, lighting, and mood, all stated with intent rather than hope. From there, we layered in negative prompts to strip out the artefacts and clutter you never wanted in the first place.

Then came the part that separates hobbyists from professionals: disciplined iteration. You change one variable at a time, note what shifts, and stop guessing.

We also looked at brand consistency, because a marketer at a Singapore SME cannot ship images that clash from post to post. Locking in your palette, framing, and reference style keeps every asset recognisably yours.

Finally, we turned all of that into reusable templates. That is your prompt library, and it is the thing that compounds. Every good prompt you save becomes a shortcut for the next campaign.

Solving prompt frustration does something quietly powerful: it gives you the confidence to reach for AI image tools daily instead of dreading them. So keep testing, keep documenting, and keep growing that library.

When you are ready to go from self-taught to guided, hands-on practice, Equinet Academy’s AI Video & Image Creation Course is the logical next step. Over two days, you take AI images from creative brief to platform-ready delivery, directing them with intentional prompt engineering, evaluating them against commercial standards, and producing a set of brand-consistent visuals you can actually use. You get structured frameworks, expert feedback, and SkillsFuture support, so the leap from experimenting to producing reliable, on-brand results happens far faster than going it alone.