How to Generate Professional Images in Helmcode: The Complete Prompting Guide

The difference between an image that looks AI-generated and one that doesn't (but is) has nothing to do with the model you use. It's the prompt you write.

Most people reach for generic, conversational prompts: "Create an image of white sneakers on a dark background." And most of the time the result feels off, or just not good enough. That's normal. This guide is here to help you level up your prompts and generate stunning images you can actually use on your websites and products.

80% of the result is decided before you hit the "Generate" button. This guide is about learning to use that 80% well.

These two images were generated from the same idea: a pair of white sneakers on a black background. The only thing that changed is the prompt we gave the model.

TL;DR (if you want the short version)

• The perfect prompt is built outside the image generator first: with any LLM (or with the chat built into your Helmcode dashboard) and the meta-prompt in this guide .
• A good prompt has 6 elements, in order . The style block outranks everything else.
• If you're after a specific style, build your own "recipe" in 2 minutes.
• To generate sets of images that hang together, freeze the style block and use references.
• If something isn't working, the Troubleshooting section has you covered.

The 3 steps to follow

This is how we generate images internally. Here's the workflow we recommend:

1. Refine your prompt outside the image generator. Don't type the prompt straight into the text box. Open your favourite LLM or the Helmcode chat, tell it your idea (or hand it a reference image of the style you're after) and ask it to build the prompt for an image generator. Below we give you the exact text so you can see how to ask for it.

2. Configure the generation. Paste the refined prompt into the image generator (or, if you're using the API, into your console or IDE) and set three things: the aspect ratio, how many variants you want, and the Enhance Prompt toggle.

As a rule of thumb we recommend leaving the toggle on, because it improves your prompt even further, but there are cases where it isn't a good idea. We cover them in their own section.

3. Generate and iterate. Sometimes you don't nail it on the first try. Open the image, check the seed in the detail panel and, if you want to test a specific change, reuse that seed to compare against the first image. If you like an image as a base, hit "Use as reference" and the model will use it as the starting point for the next one.

The meta-prompt: copy and paste

This is one of the most valuable parts of the guide. Paste this block into your favourite AI assistant, add your idea at the end (or attach a reference image) and it will hand you back a prompt that's ready for Helmcode:

Act as an expert prompt writer for image generation models (Flux and similar). Your job is to turn my idea, or the reference image I attach, into a refined, ready-to-use prompt.

RULES:
1. Always write the prompt in English, using natural descriptive language (full sentences, not comma-separated tag lists).
2. Structure the prompt in this order: subject in concrete detail, then environment/surface, then lighting, then background, then camera and lens, then a style block at the end.
3. Lighting defines 80% of the mood: decide the mood first, then choose the lighting vocabulary accordingly (soft diffused light, harsh direct sunlight with sharp shadows, moody low-key spotlight...).
4. Name the exact palette when the style calls for it ("limited palette of black, vermilion red, cream and gold").
5. Anchor the style with technique or era references to escape the generic digital look ("screenprint", "risograph", "1970s magazine", "2000s paparazzi flash", "medium format film").
6. Negate what the model tends to add ("flat colors, no gradients" for print styles; "desaturated, muted tones" if colors come out too vivid).
7. If the image must include text, put the exact text in quotes and describe the style of each text element separately.
8. If I give you a reference image: identify the medium, extract the palette as named colors, note the textures, the lighting and the composition, and build the prompt from there.

RESPONSE FORMAT:
- The final prompt in a copyable block.
- One line telling me which 2-3 terms in the prompt drive the look the most, in case I want to adjust.
- Recommended aspect ratio based on the use I describe (1:1, 16:9 or 9:16).
- If the prompt uses precise technical vocabulary (exact palettes, quoted text, effects like dithering), remind me to turn OFF "Enhance Prompt" in Helmcode; if it's a short, general prompt, tell me to leave it on.

MY IDEA / REFERENCE:
[describe what you want here, or attach the image]

A real example: your prompt and the result

I run an online store selling handmade olive oil soaps and natural skincare. I need product photos for the website's category page and for Instagram. Our brand feels artisan, warm and Mediterranean, nothing clinical or corporate. The photos should make the products look premium but approachable, like something you'd find in a nice local market. I'll be generating several products over time, so I want a style I can repeat.

Here's the prompt the meta-prompt hands back:

A single bar of handmade olive oil soap, wrapped in unbleached kraft paper tied with a thin jute twine, resting on a worn, rough-hewn wooden tabletop with visible grain and slight salt stains. The soap is a pale, matte green, unevenly cut at the edges, with a subtle marbled swirl of darker green through the center. The lighting is warm, directional, side-lit morning sun, casting long soft shadows across the table, with a gentle rim of golden light catching the twine. The background is a shallow depth-of-field view of a whitewashed plaster wall, lightly cracked, and a blurred terracotta pot with a single sprig of dried rosemary. Shot with a 50mm f/1.8 lens at close focus, shallow depth of field, slightly warm white balance. Style: analog medium format film, slightly desaturated, high grain, flat natural colors, no digital sharpening, no gradients. The texture of the wood, soap, and paper should feel tactile and dry, not glossy or plastic.

The terms driving the look: warm, directional, side-lit morning sun + casting long soft shadows + slightly desaturated. Aspect ratio: 16:9.

And here's the result:

​* The prompt output is written in English on purpose: results with an English prompt are noticeably better.

** You can turn this meta-prompt into a skill if you'd rather not hunt it down and copy-paste it every time.

Anatomy of a good image prompt

If you'd rather write your prompts by hand (or understand what the LLM is doing for you), these are the 6 pieces every good prompt should have:

• Subject. What appears in the image, described with precision: what it is, pose, expression, materials. A ceramic vase with a single olive branch works better than a nice vase.
• Environment or surface. Where the subject sits and what surrounds it: a stone pedestal, a wrinkled linen table, the wet asphalt of a street.
• Lighting. The single piece that changes the result the most. soft diffused lighting gives a clean, editorial mood; harsh direct sunlight with sharp shadows gives Mediterranean life; moody low-key spotlight gives drama and luxury. Decide the mood first, then pick the lighting words.
• Background. A flat color, a seamless studio, or a full scene.
• Camera and lens. Angle, focal length and depth of field. Shot on 85mm lens, shallow depth of field is the most reliable shortcut to a professional look.
• Style block. At the end of the prompt: medium, palette, texture, era or technique references. This is where the image's personality lives.

The star trick for working in series: when you want several images with a consistent style (a catalogue, a campaign, a feed), change only the subject sentence and leave the style block untouched. The style block is what does the magic.

Two notes on the prompt. Length has a sweet spot, between 40 and 80 words: shorter comes out generic; longer, and the model starts ignoring the tail of the prompt. And order matters: what comes first gets more attention, which is why the subject opens and the style closes. If something in the style block keeps getting lost in the results, move it up. And if you go over 80 words it's fine, just try not to pass 100.

Learn to generate your images from scratch (without copying)

If we handed you 20 ready-made prompts, the moment you wanted to create something beyond those 20 you'd be stuck. So here are the concrete steps to build your own image recipes. You do it in your favourite LLM, with the meta-prompt above.

1. Give the assistant business context. Don't just tell it what you want to see; tell it what it's for and who you are. Alongside the meta-prompt, add a line like "I'm a natural cosmetics brand, the image is for the website home" or "I'm a restaurant, this goes on the digital menu". The destination decides the aspect ratio and your brand's mood decides the light.

2. Ask for the prompt in two pieces: subject and style block. The subject is what changes in every image; the style block (medium, palette, light, texture, era or technique references) is what repeats. That separation is the recipe.

3. Adjust one term at a time. Generate, look at the result and, if something's off, change a single thing (usually the lighting vocabulary or the palette) while keeping the same seed so you can compare cleanly.

4. Save the style block once you nail it. That final sentence of the prompt is your brand recipe. From then on, every new image is just swapping the subject sentence and generating. A whole catalogue, an entire feed or all the illustrations on your site can come from the same block.

Let's take the earlier example. You tell the assistant:

I run an online store selling handmade olive oil soaps and natural skincare. I need product photos for the website's category page and for Instagram. Our brand feels artisan, warm and Mediterranean, nothing clinical or corporate. The photos should make the products look premium but approachable, like something you'd find in a nice local market. I'll be generating several products over time, so I want a style I can repeat.

And it hands you back something like this:

Subject: A single bar of handmade olive oil soap, wrapped in unbleached kraft paper tied with a thin jute twine, resting on a worn, rough-hewn wooden tabletop with visible grain and slight salt stains. The soap is a pale, matte green, unevenly cut at the edges, with a subtle marbled swirl of darker green through the center.

Style block: The lighting is warm, directional, side-lit morning sun, casting long soft shadows across the table, with a gentle rim of golden light catching the twine. The background is a shallow depth-of-field view of a whitewashed plaster wall, lightly cracked, and a blurred terracotta pot with a single sprig of dried rosemary. Shot with a 50mm f/1.8 lens at close focus, shallow depth of field, slightly warm white balance. Style: analog medium format film, slightly desaturated, high grain, flat natural colors, no digital sharpening, no gradients. The texture of the wood, soap, and paper should feel tactile and dry, not glossy or plastic.

Aspect ratio: 16:9. Enhance Prompt: Off.

Whenever you need an image of a new soap, you just change the first sentence. The style doesn't change.

Here are 3 new images generated with the same style block and a different subject. In this case, 3 new candles for my e-commerce.

Enhance Prompt: when to use it and when not to

Helmcode includes an Enhance Prompt toggle that automatically rewrites your prompt, adding detail without changing the subject. It's on by default, and for most cases that's fine.

When you generate an image, the detail panel shows you your original prompt and the enhanced prompt that actually ran. They're worth reading: it's a free prompting lesson with every generation.

When to leave it on? When your prompt is short or general, or when you want fast results without wrestling with vocabulary.

When to turn it off? When you've crafted a prompt with precise technical vocabulary: exact color palettes, quoted text that must appear verbatim, specific effects like dithering or misregistration. The rewriter can dilute the very terms that define your look. If you used the meta-prompt from this guide, the assistant itself will tell you what to do with the toggle.

If something looks off: quick fixes

The model's typical failures and how to try to fix them:

• Weird hands. The classic. Don't fight them: change the pose or the framing so the hands aren't the star (hands in pockets, holding a cup, arms crossed), or crop above them. If the hands are essential, generate 4 variants and pick.
• Illegible or invented text. Cut it down to one or two words and always put them in quotes in the prompt; if it still fails, generate the clean scene and add the typography afterwards in your editor. Past three words, the failure rate climbs fast.
• Warped faces in wide shots. The smaller the face in the frame, the worse the model resolves it. Move in (medium shot, close-up) or reduce the number of people in the scene.
• Everything has "AI face". That polished, too-perfect finish is cured with imperfection: add grain (film grain, 35mm film photography), an era (1990s snapshot, 2000s paparazzi flash) or imperfect light (slightly overexposed, harsh direct flash).
• The model ignores part of the prompt. Usually happens with long prompts: trim to the essentials and move up whatever's getting lost. And check the enhanced prompt in the detail panel, because sometimes the culprit is the rewriter and the fix is turning the toggle off.
• Colors come out too vivid. Negate it in the prompt itself: desaturated, soft muted tones. And if you're after a print style, flat colors, no gradients helps too.

Start generating

You've got the full workflow now: refine the prompt with the meta-prompt from this guide, set format and variants in Helmcode, and iterate with the seed and references until you nail it.

If you want the detailed reference for every control (aspect ratios, variants, quotas, advanced panel), it's in the Images docs. And if you use Claude daily, we have a prompting skill that automates the whole refining process.

Now it's your turn to burn some tokens and generate great images. Head to cloud.helmcode.com/images and go for it.