Midjourney’s own prompting rules are shorter than the guides written about them: keep the prompt brief, cover the elements it names, put every parameter at the end. Since V8.2 you can also write an instruction instead of a description, which is where most tutorials are still out of date.
What the official prompt pages say
The definition is one sentence: “A prompt is simply the text or phrase you use to tell Midjourney what kind of image or video you want.” The rest of the Prompt Basics page follows from it.
The loudest rule is the one tutorials tend to bury: “Short and simple prompts typically generate the best images with Midjourney… Avoid making long lists or detailed instructions; these can confuse the process.”
Midjourney shows the difference rather than arguing it. The version to avoid: “a picture of lots of blooming California poppies, make them bright, vibrant orange, and draw them in an illustrated style with colored pencils”. The preferred version: “Colored pencil illustration of bright orange California poppies”.
Two smaller tips sit under that: name a number instead of a plural, and describe what you want rather than what you don’t, since a prompt about a party “with no cake” can still put a cake on the table. The Midjourney quick-start guide covers the first grid.
The seven elements, and the trade-off behind them
Midjourney names the areas worth covering and then stops, which is what keeps the list inside one sentence.
| Element | The question the official page asks | Phrases it suggests |
|---|---|---|
| Subject | Who or what? | person, animal, character, object |
| Medium | In what form? | photo, painting, illustration, sculpture |
| Environment | Where? | indoors, outdoors, underwater, in the city |
| Lighting | What kind? | soft, ambient, overcast, neon, studio lights |
| Color | In what shades? | vibrant, muted, monochromatic, pastel |
| Mood | Feelings to evoke? | playful, calm, gloomy, energetic |
| Composition | How is it framed? | portrait, headshot, closeup, birds-eye view |
Filling all seven has a documented price: “Fewer details mean more variety, but you get less control over the outcome.” Two elements buy exploration; seven plus a parameter buy a render you can repeat.
Description or instruction: the split V8.2 introduced
The Edit Model page carries the sentence that changes how a prompt should be written: “The traditional way to write prompts in Midjourney is to describe the image or scene you are trying to create, without giving directions or instructions. However, with the Edit model, you can now provide specific instructions in your prompts.”
Four reference images, and --edit in Discord
The Edit Model accepts “up to 4 reference images” and replaced Omni Reference and Character Reference, so the --cref and --oref in older tutorials describe a route that no longer exists. On the website you attach images; in Discord you add --edit. It does not work with --tile, and its results cannot go through Remix.
The aspect ratio follows the first reference image
An edit job inherits the shape of the first image you attach, not your saved default, so --ar has to be written into the prompt when the output should differ. Adding --raw is the documented way to give the instruction more weight.
Parameters: three rules that decide whether a prompt runs
Parameters carry the settings — --ar, --stylize, --chaos, --no, --raw, --version, the speed flags — and the Parameter List reduces their formatting to three rules.
The three formatting rules
- Place parameters at the end, after the prompt text.
- Put a space between the prompt text and the dashes.
- Use no commas, periods or other punctuation inside a parameter.
Four failing prompts make the rules concrete: poppies--ar 2:3, poppies - - ar 2:3, poppies --ar 2:3, and --ar 2:3 poppies. The Midjourney API guide turns them into a check you can run before a batch of renders.
The numbers Midjourney does publish
Two values are documented to the digit. Stylize is 100 by default and adjustable “anywhere between 0 and 1000”, so a low value keeps the prompt literal and a high one lets Midjourney interpret it. Aspect ratio defaults to 1:1, and --ar cannot contain decimals — “Use 139:100 instead of 1.39:1.”
Worked example: two prompts, two styles
The two cases below are the same subject written both ways, and the difference in the output follows the difference in the writing.
Case one: a descriptive text prompt, no image attached
a bookshop window at closing time, rain on the glass, warm interior light, one figure inside shelving a stack, muted palette, calm mood, 35mm documentary feel --ar 16:9Four variations come back together, and because the prompt is short the default style fills the gaps: the light stays warm in all four, the figure stays inside, and the framing wanders — one frame pulls back to the whole shopfront, another crops to the shelves. “Rain on the glass” and “warm interior light” do the visible work; drop either and the palette flattens. A standard render is documented at 0.8 GPU minutes.
Case two: an instruction prompt with one reference image
make this girl into a real character --ar 2:3This is the official example from the Edit Model page, attached to a drawn reference image. The published before-and-after keeps the pose, the clothing and the framing while replacing the drawing with a photographic character: the instruction changes the treatment, and the reference image holds everything the instruction leaves out. The same page turns one reference into two viewpoints with “let’s see this image from the front” and “let’s see this image from behind”. Edit Model jobs are documented at 1 GPU minute in standard resolution, 2.3 in HD.
What Midjourney has not published about prompting
Every gap on this list gets filled in with a number somewhere online.
- A maximum prompt length in words, characters or tokens.
- A range for
--chaos, or an order for the seven elements: both are named without published rules. - An official prompt formula. The seven elements are a checklist, not a template, and every structure diagram circulating is somebody’s summary of it.
- A rule for how an instruction and a description behave when mixed in one Edit Model prompt.
- A “last updated” date. The Midjourney topic hub collects the version facts that do carry dates.
Frequently asked questions
Is there an official Midjourney prompt formula? No. The pages publish a checklist and a length preference, and they never rank the items.
Can I write instructions instead of descriptions? With the Edit Model, yes. Without an attached image an instruction has nothing to act on.
What replaced --cref and --oref? The Edit Model. The parameter list now marks Omni Reference as replaced by it in V8.X.
Does a single word or an emoji really work? The official page says prompts “can be simple—even a single word or emoji works”, and that short prompts let the default style fill the gaps.
Why does a longer prompt give flatter results? Fewer details buy variety, more buy control, so a long prompt narrows what Midjourney explores. The quick-start guide shows that on a first render.