Midjourney Prompt Playbook

Updated: 2026-08-29. Prompt syntax, parameter names, and defaults follow midjourney.com and official docs that day. Midjourney has no official public API; third-party entries ≠ official service.

Introduction

Searches for “Midjourney prompt,” “Midjourney parameters,” or “how to write Midjourney prompts” usually lose quality by stacking adjectives. The gap that matters is writing prompts as checkable art briefs: use, canvas, subject, composition, light, material, bans—then hang --ar, --stylize, --weird, and style/character refs when needed. This guide gives a fixed structure, three paste-ready templates, and stable patterns for parameters and negatives so you spend less time on “pretty grid, undeliverable final.” Pair the loop with Getting started.

What this guide solves

• Turn empty prompts into reusable generate specs with a segmented structure
• Get three copy-ready templates: product still life, poster/social, local edit
• Use --ar / --stylize / --weird correctly (official that day)
• Understand what --sref / --cref / --oref-class refs actually control
• Control style drift with negatives and single-variable iteration

Write prompts as art briefs

Suggested order (you can compress, but don’t scramble the logic):

1. Use & canvas: hero / detail page / poster; aspect (UI control or --ar)
2. Subject: what it is, count, pose, must-keep brand or product traits
3. Composition & camera: center / thirds, whitespace, angle and focal feel
4. Light & palette: key-light direction, contrast, brand or banned colors
5. Material & style: photo / 3D / illustration / flat infographic; anime track → Niji 7
6. On-image text: prefer post; if required, give short exact strings and placement
7. Bans: watermarks, extra limbs, cluttered backgrounds, unauthorized logos
8. Parameter suffix: --ar etc. (syntax per official docs that day)

Empty: “draw a nice café” → hard to accept.
Spec: “16:9 blog hero; latte with latte art slightly left of center; shallow 50mm; window light; warm cream + deep green; photo-real product; no on-image text; no logos or watermarks --ar 16:9” → iterable.

English, Chinese, or mixed?

Both languages can generate. Practice tips:

• Team reviews use your shared language on the acceptance sheet; prompts may mix
• Proper nouns, materials, and lens words: write them in whichever language you nail more reliably
• Don’t stack synonym fights across languages (e.g. “photorealistic + ultra-realistic + photo-grade” three times)

Turning “inspiration” into a deliverable spec

Many people open Midjourney with one line of vibe: “cyber rain-night café.” Translate to a spec before you generate:

Use: 16:9 blog hero with left whitespace.
Subject: one hot americano on a wood window table; neon blur outside; only this table in frame indoors.
Light: cool neon outside + warm desk lamp; strong contrast without blowouts.
Style: editorial photography; mild film grain optional.
Ban: readable signage text, watermarks, extra face close-ups, sci-fi mechs.
--ar 16:9

Inspiration sets direction; the spec sets acceptance. Without step two, a beautiful grid still fails handoff. Once the spec lives in a team template, new people reuse the same structure instead of inventing spoken dialects.

Word order & weight habits

Midjourney usually weights concrete nouns earlier (exact behavior follows the model that day). Habits:

1. Front: hard constraints — use, subject identity, count, aspect intent.
2. Middle: picture language — composition, light, material, one style line.
3. Back: bans & parameters — negative list and --ar suffix.

If “costume detail written at the end always drops,” move costume layers into the subject block. That beats repeating the same adjective three times.

Three copy-ready templates

Product still life

Use: ecommerce detail main image, 1:1.
Subject: [full product name & material], fully visible, mild 3/4 angle.
Background: clean [color] seamless, soft contact shadow on the ground.
Light: soft top-side light; highlights not blown.
Style: commercial product photography, real material texture.
Text: none.
Ban: hands, damaged packaging, watermarks, extra props stealing focus.
--ar 1:1

Poster / social cover

Use: social cover, 16:9.
Theme: [one-sentence theme].
Composition: large left whitespace for title; subject on the right.
Subject: [person or object], mood [keyword].
Style: [flat illustration / editorial photo / 3D render].
On-image text: none (title overlay in post). If must generate, short title only: "[short line]", top-left safe zone.
Palette: [primary] + [secondary].
Ban: legal microcopy, QR codes, extra logos, watermarks.
--ar 16:9

Local edit (Vary / multi-round)

Keep the previous camera, pose, palette, and background unchanged.
Change only: [single variable, e.g. swap the cup for a closed notebook].
Do not redraw the whole composition; do not change subject identity or light direction.
Do not add text or watermarks.

One variable per round—or you cannot tell which sentence broke the frame. Grid and Upscale: Getting started; model choice: V8.2 and Niji 7.

Migrating templates into a project library

Personal folders often hold dozens of “once worked” prompts nobody else can use. Minimum migration bar:

1. Strip one-off subject names back to [placeholders].
2. Note the model track and aspect used that day.
3. Attach one success thumbnail and one failure note (if any).
4. Delete expired parameter numbers; write “start from official defaults that day, then tune.”

Then templates are assets—not archaeology.

Character sheet (Niji-track short form)

Use: full-body character sheet, solid background for cutout.
Subject: [hair & color], [costume layers], [signature accessory], expression [word], pose [word].
Style: clean line art, [flat/thick paint], anime illustration.
Ban: real-skin pores, photo noise, cropped feet, extra fingers, on-image text.
--ar 2:3

Full Niji workflow: Niji 7.

Template discipline

1. Fill brackets before improvising: generating with empty [product name] is still an empty prompt.
2. One template type per job: don’t mash product still life and poster into one mega paragraph.
3. On success, save a project template: freeze brand colors, banned logos, fixed aspect—part of the style card.
4. On failure, return to the template—don’t add a tenth adjective.

Parameters: --ar, --stylize, --weird

Below is conceptual; exact spellings, abbreviations, defaults, and limits follow official docs that day. Old blog numbers may be dead.

--ar (aspect)

Common use	Aspect intent
Blog / landscape hero	16:9
Ecommerce main / avatar	1:1
Vertical poster / phone	9:16 or 2:3
Wallpaper feel	16:9 or wider (if supported that day)

Lock channel crop first, then write content. If Web already set aspect via controls, avoid fighting a suffix—follow entry behavior that day.

--stylize / --s (stylization)

Controls how hard the model “follows its own taste” (conceptually).

• Lower: more obedient, closer to the brief—good for product specs and character sheets
• Higher: prettier, easier to drift—good for moodboards and concept exploration
• Beginners: start default or mid-low; when delivery is unstable, lower stylize before blaming the model

--weird (weirdness)

Raises odds of unconventional composition and odd combinations (conceptually).

• Learning and commercial finals: most jobs can skip it
• Concept art, surreal posters: small probes; accept higher reject rates
• Combined with hard product specs: acceptance often fails

Other common parameters (touch lightly)

Version, quality, seed, stop values—and their names—change over eras. Principles:

1. Tune one parameter at a time against the same prompt
2. Log working combos on a team parameter card
3. Official docs that day win—this article is not a parameter-manual reprint

How to log parameter A/B tests

Same prompt skeleton; change one suffix only. Minimum table:

Date	Entry	Model	Prompt summary	Parameter delta	Grid feel	Adopt as default?
2026-08-29	Official Web	(name that day)	lamp hero	stylize default→lower	closer to brief	yes

Two weeks later, your own table decides team defaults—not a forwarded “mystery combo” screenshot. Prices and plans still follow pages that day; don’t hard-code unit prices in the table.

Reference class: --sref, --cref, --oref

Names and availability follow official that day. Mental model:

Type	Common flag	Controls	Writing tip
Style ref	--sref	Palette, brush, overall vibe	“Want style, not composition copy”
Character ref	--cref etc.	Same person recognizable	Fixed identity sentence; change only scene/action
Object ref	--oref etc.	Product/prop silhouette	“Geometry follows the reference”

Practice rules

1. Qualified master first — bad master composition amplifies the wrong vibe.
2. One job per reference — geometry / palette / identity written apart.
3. Conservative weights first — too strong locks; too weak does nothing; tune per docs that day.
4. Track match — photo masters on Niji characters (or reverse) feel wrong.
5. Rights — unlicensed images are not commercial evidence.

Style-card capture: Styles & scenes.

Reference prompt sketch

[subject & scene specs…]
Style ref: match the attached style image’s line weight and palette; do not copy its pose or background clutter.
Character ref: keep the same person identity and hair; this round only changes “standing by a window in side light.”

Negatives & diagnostic table

Writable “don’ts” beat empty “high quality.”

Symptom	Common cause	Next round
Style jumps	Too many conflicting adjectives	Cut to 1 primary style + 1 reference material
Subject warps	Weak constraints or too many changes	Lock camera/pose; change one object
Soft / wrong text	Too much or too small copy	Shorten or no-text + post type
Brand mismatch	Missing refs or must-keep traits	Upload refs; “keep pack geometry and color blocks”
Dirty background	No bans written	“Clean background, no clutter, no watermarks”
Hands/feet collapse	Integrity never called out	“Both hands five fingers clear; both feet fully in frame”
Wrong aspect	No --ar / control	Bind use to aspect before generating

Paste-ready ban line

Ban: on-image text, watermarks, borders, QR codes, unauthorized logos, extra limbs, fused fingers, cropped subject, clutter props stealing focus.

Embed the diagnostic table in reviews

Weekly reviews shouldn’t stop at “this looks bad.” Speakers point to one diagnostic row: style jump, warp, text, brand miss, dirty background, limbs, or aspect. Fix the next round accordingly—meetings get shorter. If the issue is “wrong track,” switch to Niji 7 or V8.2 instead of adding adjectives.

On-image text that stays stable

• Keep copy short; quote exact strings
• State language and rough placement (top-left / bottom safe zone)
• Zero-error spelling or complex trademark glyphs: no-text first, then layout tools
• Legal microcopy, dense tables, QR codes: don’t expect a perfect one-shot

Recommended text workflow

1. Default: no-text generate → design tool overlays titles / logos / disclaimers.
2. Short-title path: quote exact short line → separate round for text only → human proof.
3. Forbidden path: dump long paragraphs, mixed scripts, and tables into one prompt and expect every grid cell to be perfect.

Brand fonts, registered marks, nutrition tables: almost always safer in post. Write that into the style card to cut “AI text wall” rework. Scene capture: Styles & scenes.

Completion checklist

• Prompt includes use, aspect, subject, light, style, bans
• Parameter suffix doesn’t fight UI controls; spelling follows official that day
• This round changed one variable (prefer either prompt or parameter—not both)
• Reference rights and control scope stated
• On-image text proofed (or post-type decided)
• Limbs and artifacts checked at thumbnail and target size
• Logged model track (general / Niji), entry, and prompt summary

Quick access

• China-friendly (third-party): Midjourney V7
• Mirror studio (third-party): Text-to-image studio
• Official: www.midjourney.com

Third-party accounts, logs, and routing may differ; don’t casually upload sensitive assets. More paths: China access. Again: no official public API.

FAQ

Are longer prompts always better?

No. Long, mutually exclusive adjectives lower controllability. Hard constraints first, then style—beats twenty near-synonyms.

Does maxing --stylize make it prettier?

It can look flashier while obedience drops. When delivery is unstable, lower stylize before stacking “8K masterpiece.”

Parameters at the start or end of the prompt?

Community habit is usually at the end; follow official docs and your entry’s parser that day. Fix one placement per task so comparisons stay fair.

Why doesn’t --sref match?

Check: is the master good, is ref duty written clearly, is weight too high/low, does the model track match? Swapping masters often beats stacking adjectives.

Must I prompt only in English?

No. Either language works; don’t fight synonyms across languages in one line. When teams accept in a shared language, hard constraints in that language usually help collaboration and retros.

Prices and quotas?

Follow official or third-party pages that day; this article does not hard-code prices.

Official resources

• midjourney.com
• Official parameter and prompt notes (in-site help/docs entry that day)

Further reading

• Midjourney getting started: first image
• Midjourney styles & scenes
• Midjourney V8.2 full guide
• Midjourney Niji 7 selection & practice
• Midjourney Guides Hub
• What is Midjourney? Family map

Summary

Midjourney prompts win as art brief + single-variable iteration + parameters per official that day: use and aspect first, subject and bans checkable, twist only one of --ar / --stylize / --weird at a time, and declare control scope on every reference. Park text and complex marks in post; freeze wins as templates. That beats chasing “mystery spells.” Third-party entries need their own terms review—and do not assume an official public API exists.