midjourney-prompt-engineering

Midjourney Prompt Learning System

A skill that knows Midjourney. The foundation is a structured understanding of Midjourney V7 built from the official documentation — every parameter, prompt syntax rule, reference system, and style code mechanic. On top of that, a learning loop: each session extracts patterns from what worked and what didn't, building a knowledge base of craft that improves first-attempt quality over time.

Architecture

You are a multimodal reasoning model. You don't need pipelines — you ARE the visual critic, gap analyzer, and prompt rewriter. You analyze MJ output images directly, score dimensions, identify gaps, and rewrite prompts.

The one thing you can't do natively is remember across sessions. That's what the persistent layer provides — the database, patterns, and evidence tracking.

Knowledge Foundation (ships with the skill)

File	What It Contains	Source
knowledge/v7-parameters.md	Every V7 parameter, prompt structure rules, breaking changes from V6	Official docs
knowledge/translation-tables.md	Visual quality → prompt keyword mappings (lighting, mood, material, color, composition)	Official docs + tested refinements
knowledge/official-docs.md	Documentation map linking each MJ feature to its official page URL	docs.midjourney.com
knowledge/failure-modes.md	Diagnostic framework for common MJ failure patterns	Session-learned, evidence-backed
knowledge/learned-patterns.md	Auto-generated pattern summaries (grows through use)	Extracted from sessions
knowledge/keyword-effectiveness.md	Keyword effectiveness rankings (grows through use)	Extracted from sessions

The static files (v7-parameters, translation-tables, official-docs) are the skill's baseline knowledge — what a skilled MJ user would know from reading the documentation carefully. The dynamic files (failure-modes, learned-patterns, keyword-effectiveness) are populated through real sessions and grow over time.

Module Dependencies

Module	Purpose	Required MCP
Core rules (core-*)	Reference analysis, prompt construction, scoring, iteration	None
Learning rules (learn-*)	Pattern lifecycle, reflection, keyword tracking	sqlite-simple
Automation rules (auto-*)	Browser automation for midjourney.com	playwright

Core only (manual): Load core-* rules. Copy prompts to MJ manually. Core + Learning: Add learn-* rules + sqlite MCP. Patterns persist across sessions. Full system: Add auto-* rules + playwright MCP. Hands-free iteration.

# SQLite (for learning rules)
claude mcp add sqlite-simple -- npx @anthropic-ai/sqlite-simple-mcp mydatabase.db

# Playwright (for automation rules)
claude mcp add playwright -- npx @playwright/mcp@latest --headed

# Initialize the database
sqlite3 mydatabase.db < schema.sql

Rules Quick Reference

Rule	What It Covers
core-reference-analysis	7-element visual framework, vocabulary translation
core-prompt-construction	V7 prompt structure, keyword practices, knowledge application
core-research-phase	Coverage assessment, community research workflow
core-assessment-scoring	7-dimension scoring, confidence flags, agent limitations
core-iteration-framework	Gap analysis, action decisions, aspect-first approach
learn-data-model	Database schema, session structure, ID generation
learn-pattern-lifecycle	Confidence graduation, decay, knowledge generation
learn-reflection	Session lifecycle, automatic reflection, contrastive analysis
auto-core-workflows	Prompt submission, smart polling, batch capture, animation
auto-reference-patterns	Selector strategy, error handling, image analysis

Scoring

All iterations scored on 7 dimensions: subject, lighting, color, mood, composition, material, spatial. All 7 always scored (1.0 for "not applicable"). Scores are preliminary until user-validated. See rules/core-assessment-scoring.md.

Commands

Command	Purpose
/new-session	Start a session with full knowledge application
/log-iteration	Log a generation attempt with scoring and gap analysis
/reflect	Cross-session pattern analysis and knowledge extraction
/research [focus]	Research community techniques for a challenge
/show-knowledge [category]	Display learned patterns
/apply-knowledge <desc>	Pattern-informed prompt for a description
/discover-styles	Browse and catalog MJ style codes
/validate-pattern [id]	Mark pattern as validated or contradicted
/forget-pattern [id]	Deactivate a pattern

Key Principle

Every pattern must have logged evidence. The system learns from real iteration data, not assumptions. Confidence levels (low/medium/high) reflect how many times a pattern has been tested and its success rate.

Full Reference

For the complete compiled reference combining all rules, see AGENTS.md.