A freelance designer I know burned an entire Tuesday typing “cool retro poster” into Midjourney. Forty images later she still had nothing she could send a client. The next morning she changed one habit: subject first, then medium, then light, then framing. Three attempts and she had the artwork.
That habit is the answer. To write effective Midjourney prompts, name your subject in concrete nouns, state the medium or art style, add setting, lighting and color, then steer the output with parameters such as aspect ratio. Keep the whole thing to roughly 60 words, separate ideas with commas, and treat each generation as a draft you revise rather than a lottery ticket.
This guide is for people who can already run the tool but keep getting images that look almost right. You get a formula, a weak-versus-strong comparison table, a costed worked example, a glossary, and a framework for picking a prompt style.
The Four-Part Prompt Formula
Nearly every prompt that works can be broken into four blocks. Write them in this order.
- Subject. Who or what is in the frame, plus what they are doing. “Three enormous mice eating a house made of cheese” beats “mice.”
- Medium and style. Oil painting, charcoal sketch, film still, 3D render, product photography, cyanotype.
- Setting, light and color. Time of day, weather, light quality, palette. “Overcast harbour, cold blue-grey light.”
- Controls. Aspect ratio, exclusions, reference images, model version.
Why the order changes the picture
Word position carries weight. Zapier’s prompt guide shows a case where a close-up failed until the writer moved the fire from the subject slot to the background slot, and the composition finally matched the request, per its AI art prompt walkthrough. Front-load what must be sharp. Push context to the back.
The 20-second sanity check
Read your prompt and ask three questions. Could a human illustrator draw this without asking me anything? Have I said what kind of picture it is, not just what is in it? Have I named the light? If you answer no to any of them, you are about to waste a generation.
Understanding Midjourney Prompts and What Makes Them Work
A Midjourney prompt is a short text description, optionally combined with reference images and flag-style parameters, that the model converts into four image variations. It is not a search query. Nothing is being retrieved. The model is producing pixels that statistically match your description, which is why vague wording produces generic output.
Length: the sweet spot most people miss
Longer is not automatically better. Midjourney responds well to prompts of around 60 words, and piling on extra clauses can confuse any image model, according to Zapier’s guidance on writing art prompts. Commas help. They break your request into chunks the model can weigh separately, and they make your own prompt easier to edit later.
Vocabulary is the real skill
Prompting an image model rewards a specific kind of literacy. IBM notes that people writing prompts for image generators benefit from knowing art history, photography and film terminology, in its overview of prompt engineering. “Shallow depth of field” does something. “Nice photo” does not. Build a personal word bank of lenses, movements, lighting setups and camera angles.
Words versus parameters
Some things belong in prose, some belong in flags. Shape, exclusions and reference weighting are handled by parameters. Zapier’s Midjourney walkthrough documents several: --no suppresses an element you keep getting, --ar sets a non-square aspect ratio, --sref points at a style reference with --sw weighting between 0 and 1,000, and --cref points at a character reference with --cw weighting between 0 and 100. Both weights default to 100.
Zero-shot, few-shot, and why it matters here
IBM describes zero-shot prompting as asking for something the model was never explicitly trained to do, and few-shot prompting as supplying examples so it can infer what you want, in the same prompt engineering explainer. Image generation has a visual equivalent. A bare text prompt is zero-shot. Attaching a style or character reference is your few-shot example.
Tips for Writing Effective Midjourney Prompts
These are the moves that reliably lift image quality, ordered roughly by payoff.
- Replace adjectives with nouns. “Gigantic” carries more than “big, large, huge” stacked together.
- Name the medium every single time. It is the cheapest way to escape the default house style.
- Describe light explicitly: golden hour, hard chiaroscuro, overcast softbox, neon spill.
- Give a color scheme rather than individual object colors when you want a coherent look.
- Add camera language for realism: focal length, shot type, film stock.
- Use
--noto remove a stubborn element instead of writing “without.” - Set aspect ratio before you fall in love with a square composition you cannot use.
- Change one variable per attempt so you learn what actually moved the needle.
- Keep a text file of prompts that worked, tagged by use case.
- Upload a reference image when the brief has a fixed look you must match.
A reusable template
Paste this and fill the brackets: [subject doing something], [medium/style], [setting], [light], [color palette], [composition] --ar [ratio] --no [unwanted element]. It looks mechanical because it is. Mechanical beats inspired when you need to create a usable image in ten minutes.
Reference images earn their keep
You can attach an image and tell Midjourney to treat it as the base of the prompt, as a style source, or as a character source, and you can combine those roles, per Zapier’s hands-on Midjourney guide. For brand work this is the difference between “roughly on-brand” and “usable.” Other platforms expose comparable controls in their own way; OpenAI, for example, publishes its image generation options in its API image guide.
Common Mistakes to Avoid
Most bad output traces back to a short list of habits.
- Adjective spam. Ten intensifiers dilute each other. Pick the strongest word and stop.
- Silent medium. No medium means the model picks one, and it will pick its favourite.
- Contradictions. “Minimalist, highly ornate baroque” gives the model an impossible brief.
- Negation in prose. “No text” often produces text. Use the exclusion parameter instead.
- Rewriting everything after one miss. You lose the signal about what worked.
- Ignoring aspect ratio until you need a banner and only have squares.
- Copying a viral prompt wholesale without swapping the subject to yours.
- Judging a prompt on one grid. Four images is a small sample of a random process.
- Forgetting your images may be public on default plans, which matters for unannounced client work.
Why stacking synonyms backfires
Every word competes for attention inside the same budget. Adding “big, huge, massive” spends three slots on one idea and starves the details that actually distinguish your image, such as texture or lens. Trim ruthlessly. If a word does not change the picture, delete it.
When to reroll instead of rewrite
If three of four images are close, rerun the same prompt. Generation is stochastic, so identical input produces different results. If all four are wrong in the same way, the prompt is at fault and no amount of rerolling will help.
Case Studies: Effective vs. Ineffective Prompts
The table below shows the same four briefs written badly and well, plus what changes in the output.
| Brief | Weak prompt | Stronger prompt | What improves |
|---|---|---|---|
| Café menu header | coffee shop picture | overhead flat lay of a flat white and two croissants on a scratched oak table, natural window light, warm cream and umber palette, food photography, 50mm –ar 3:1 | Usable crop, consistent palette, no random signage |
| Fantasy book cover | cool dragon art | a small green dragon asleep on a stack of library books, oil painting, candlelit reading room, deep amber and ink black, three-quarter view –ar 2:3 | Correct portrait ratio, controlled mood, single subject |
| Product hero | headphones on white | matte black over-ear headphones on a pale concrete plinth, studio product photography, single softbox from the left, cool grey background, shallow depth of field | Clean edges, believable shadow, brand-neutral tone |
| Editorial illustration | drawing about remote work | flat vector illustration of a person working at a kitchen table with a cat on the laptop, limited four-color palette, thick outlines, midday light | Consistent style across a series, printable |
Worked example: a ceramics brand’s spring launch
A two-person ceramics studio needed nine lifestyle images for a spring collection. Their first session used prompts like “nice pottery photo” and produced 48 images across 12 generations, of which one was usable. Cost of that session on a plan that includes roughly 200 fast generations for about ten dollars a month, as reported in Zapier’s Midjourney pricing rundown, was around 6% of the monthly allowance for a single keeper.
The second session used the template above. They fixed the subject line (“a hand-thrown speckled stoneware mug”), locked the medium (“editorial product photography”), locked the light (“soft north-facing window light”), locked the palette (“warm sand, sage, off-white”), and set --ar 4:5 for their feed. Nine generations, 36 images, 11 usable. That is a jump from roughly 2% to 31% hit rate, and it cut the session from three hours to about forty minutes. Plans and generation limits change, so check the current tiers on Midjourney’s own site before you budget.
A small glossary worth memorising
- Aspect ratio: the shape of the frame, controlled with
--ar. - Style reference: an image used to transfer look and feel, not content.
- Character reference: an image used to keep a person’s appearance consistent.
- Inpainting: editing a selected region of an existing image.
- Outpainting: extending the canvas beyond the original edges.
- Reroll: running the identical prompt again for a fresh set of variations.
- Upscale: increasing output resolution, sometimes adding detail.
Product, Course, App and Platform Experience
Midjourney now runs as a web app as well as through Discord, and you sign in with a Google or Discord account, per Zapier’s walkthrough of the app. The web version puts creation, organisation and the community feed in one sidebar, which makes prompt iteration far less awkward than typing slash commands into a chat channel. Paid access is required; the same guide notes plans beginning near ten dollars a month and rising to a top tier around one hundred and twenty. Treat those figures as a starting point and confirm today’s plan details on the vendor’s own pricing page, since AI tools revise tiers often.
The editing surface matters as much as generation. Inpainting and outpainting let you fix one hand or widen a background without regenerating the whole image, which is usually faster than perfecting a prompt. Competing generators such as DALL·E, Firefly and Stable Diffusion make different tradeoffs, and Zapier maintains a broader comparison of image generators if you are still choosing.
If you would rather learn the underlying craft than memorise flags, explore Coursiv AI lessons for structured practice with generative tools.
Decision Framework: What to Know Before Deciding on a Prompt Style
Pick your approach by asking these questions in order.
- Is the output going to a client or to yourself? Client work needs locked ratio, palette and references from the first prompt. Personal work can stay loose.
- Do you need one image or a consistent series? A series demands a template plus a style reference, not fresh wording each time.
- Is the look already defined? If a brand guide exists, lead with a reference image and keep the text short.
- How much randomness do you want? Poetic, abstract wording gives surprise. Concrete, literal wording gives predictability. Choose deliberately.
- Will this image be public? Default plans publish generations, which can matter for unannounced products.
- Can you afford the attempts? Generation budgets are finite, so a template that lands in three tries is cheaper than inspiration that takes thirty.
Honest caveats
Prompting is not deterministic. Two identical prompts give different images by design, so no formula guarantees a specific result. Model versions change behaviour, meaning a prompt that sang last quarter can fall flat after an update. Legal ground is still shifting on AI-generated imagery, so check the platform’s current terms before commercial use rather than trusting a summary you read anywhere, including here.
Next steps
Take one image you actually need this week. Write it three ways: bare subject, full four-part formula, and formula plus a reference image. Compare the grids side by side. That single exercise teaches more than any prompt list, because it shows you what your own eye responds to.
Two more reads and you have the full picture: how to use dall e to generate images and how to use ai to edit photos.
Frequently asked questions
What is a Midjourney prompt?
How long should my prompt be?
How do I stop something appearing in my image?
--no placed at the end of a prompt to suppress an unwanted element. Writing “without trees” often produces trees.