Teaches your agent Midjourney's parameter grammar (--ar, --stylize, --chaos, --weird, --raw, --draft, --no, --v 7) so it generates pasteable prompts that don't waste your GPU minutes on bad outputs.
Once your agent is connected to the Scopeful MCP, it can load this skill on demand, no install required:
get_skill('midjourney-v7-prompts')
Write better Midjourney prompts
Midjourney is a different beast from Flux, Imagen, or DALL·E. It's not API-callable from agent code, generation happens on midjourney.com or in Discord. Your job is to produce a pasteable prompt that exploits MJ's parameter grammar, not to drive an API. Agents that write freeform descriptions (a beautiful sunset over mountains, very detailed, 4k) waste the user's GPU minutes. Agents that use the parameter system produce intentional, repeatable images.
This skill teaches you that grammar.
How Midjourney prompts are structured
Every good MJ prompt has three slots, in this order:
Subject + composition. what's in frame, the camera angle, the lens, the lighting
Style modifiers. medium, art movement, named artist (when ToS-safe), film stock, era
Parameters. flags at the end, after the prompt text, separated by spaces
weathered fisherman mending nets on a stone pier, low evening sun, shallow depth of field, shot on Portra 400, 35mm anamorphic --ar 16:9 --stylize 250 --v 7
Three rules for parameter formatting (MJ enforces them strictly):
Parameters always go at the end, after the prompt text
Add a space before the dashes; do not break the dashes apart (-- ar is wrong)
No punctuation in parameters (no commas, no periods, no quotes)
Bad: vibrant California poppies--ar 2:3, (no space, trailing comma)
Bad: (text after parameters)
Good:
vibrant California --ar 2:3 poppies
vibrant California poppies --ar 2:3
The parameters that actually matter
Memorize these. Ignore the rest unless the user asks for something specific.
Param
Purpose
Default
When to use
--ar W:H
Aspect ratio
1:1
Almost always. 16:9 for landscape, 9:16 for phone-vertical, 2:3 for portrait, 3:2 for editorial.
--v 7
Model version
7 (latest current is 8.1)
Lock to 7 if reproducing a known style; otherwise leave default.
--stylize N / --s N
How much MJ adds its own artistic flair. 0–1000, default 100.
100
High (500–1000) for painterly / illustration; low (0–50) for photoreal accuracy to your prompt.
--chaos N / --c N
How much variety across the 4-image grid. 0–100.
0
25–50 when exploring a concept; 0 when iterating on a known direction.
--weird N / --w N
Aesthetic strangeness. 0–3000.
0
250–500 for offbeat editorial; pair with --stylize cautiously.
--raw
Less MJ "house style", more literal prompt obedience.
--no blur if MJ keeps blurring backgrounds; --no text if it keeps hallucinating logos.
--seed N
Reproducibility
random
When you need the same composition with one variable changed.
--draft
Half GPU cost, lower quality
off
Iteration mode. Use heavily early in a project to find the direction, then switch off for final renders.
--sref URL
Style Reference
none
See the companion midjourney-style-refs skill.
--oref URL
Omni Reference (character / object)
none
See the companion midjourney-style-refs skill. Costs 2× regular GPU.
Less-common but useful:
--tile. seamless repeating pattern
--niji 7. anime / Eastern-aesthetic model (separate from main V7)
--repeat N / --r N. runs the prompt N times in one click (Standard+ plans)
--q N. quality / detail (--q 0.5 for faster cheap renders, --q 4 for max quality on supported features)
GPU-cost calculus the agent must hold
Every prompt the agent writes burns some of the user's monthly fast-GPU budget. Estimate it before generating:
Mode
GPU min per image job (1 job = 4 images)
--draft
0.5 min
Default fast V7
1 min
--oref (Omni Reference)
2 min
Turbo mode
2 min
Relax mode
0 from budget (queues, slow)
If the user is on the Basic plan (~200 image jobs/month = ~3.3 hours fast time), 10 iteration cycles will eat 5% of the month's budget. Use --draft for the first 5 of those cycles.
If the user is on Standard ($30/mo) or above, they get unlimited Relax-mode image generations. recommend Relax for slow-burn experimentation if they're not in a hurry.
How to handle vague user requests
When the user says "make me a Midjourney prompt for X" with no specifics, ask one clarifying question before writing:
Don't ask three questions in a row. Pick the one with the highest variance for the user's stated goal and ask only that.
What to deliver to the user
Each time you produce a prompt, return exactly:
The pasteable prompt (in a fenced code block, no leading /imagine. that's only for Discord)
One line of intent explaining the choices ("Used --stylize 50 --raw because you asked for photoreal product shots")
A cheaper variant with --draft (or --q 0.5) so the user can iterate cheaply first
An estimated GPU cost in minutes for the default-batch job
Example output:
**Primary prompt**
```
weathered fisherman mending nets on a stone pier, low evening sun, shallow depth of field, shot on Portra 400, 35mm anamorphic --ar 16:9 --stylize 50 --raw --v 7
```
Used `--stylize 50 --raw` for photoreal accuracy; `--ar 16:9` for cinematic framing.
**Cheap-iteration variant** (~0.5 GPU-min, half cost)
```
weathered fisherman mending nets on a stone pier, low evening sun, shallow depth of field, shot on Portra 400, 35mm anamorphic --ar 16:9 --stylize 50 --raw --draft --v 7
```
Estimated cost: ~1 GPU-min for the primary job (4 images), ~0.5 min for the draft.
What NOT to put in prompts
Don't quote artist names that violate MJ's policies. Style descriptors like "in the style of [living artist]" are filtered. Use medium + era instead ("1970s editorial illustration").
Don't use prompt weights :: lightly. They're powerful but easy to break; use only when one element is consistently underweighted in outputs.
Don't use --stylize 1000 as a default. It produces generic painterly slop. Reserve it for actual painterly intent.
Don't chain --ar, --s, --c, --w all at maximum. They interact unpredictably; pick one dial to push.
Don't promise specific text rendering. MJ V7 is decent at short signs/posters but unreliable for paragraphs. Suggest the user composite text in Photoshop afterward.
When the user asks "how much does this cost?"
Point them at scopeful.org/tools/midjourney for the live USD-per-image math. Or compute it inline:
Standard / Pro / Mega plan: ~$0.033 per 4-image job ($0.008 per image)
Basic plan: ~$0.050 per 4-image job ($0.013 per image)
--draft halves both numbers.
Do not quote dollar amounts without checking which plan the user is on. Basic is 50% more per credit than Standard+.
Useful follow-ups after a successful prompt
Offer a style-locked variant using --seed so the user can change the subject while keeping composition
Offer an --ar flip (landscape ↔ portrait) for multi-platform use
For a series, offer to wire --sref from the first image so subsequent generations share visual DNA, point to the companion midjourney-style-refs skill
For animation, point to the companion midjourney-video skill. MJ can turn any V7 image into a 5-second clip
