미드저니 프롬프트 엔지니어링

# Project Status## What This IsA Claude Code skill for Midjourney V7 prompt engineering. Iterates on prompts, scores results on 7 dimensions, extracts patterns from successes/failures, and accumulates craft knowledge across sessions.## Architecture```SKILL.md          — Skill definition (entry point, 134 lines)AGENTS.md         — Compiled rules (generated from rules/, 1058 lines)rules/            — 10 individual rule files + template/sections  core-*          — Prompt construction, analysis, scoring, iteration (no deps)  learn-*         — Pattern lifecycle, reflection, data model (needs sqlite MCP)  auto-*          — Browser automation workflows (needs playwright MCP)knowledge/        — V7 parameters, failure modes, translation tables, templatesscripts/build.sh  — Compiles rules/ → AGENTS.mdschema.sql        — Database initializationmydatabase.db     — Session/pattern/keyword data (sqlite)```## Context Loading Guide- **Quick orientation**: Read this file (STATUS.md)- **Architecture + workflow**: Read SKILL.md (134 lines)- **Specific rule details**: Read individual `rules/<name>.md` files as needed- **Never read AGENTS.md + rules/ together** — AGENTS.md is a compiled copy of rules/- **Deep review**: Use an Explore subagent to read everything in its own context## Database State- 11 sessions (9 success, 2 abandoned), 62 iterations, 77 patterns, 103 tracked keywords- Pattern confidence: 4 high, 39 medium, 34 low- 11 normalized categories (merged from 20 — see learn-pattern-lifecycle.md)- Sessions span: photographic, abstract graphics, complex artistic styles, style codes## Recent Improvements- Multi-reference image support (composite analysis from multiple style exemplars)- Style code knowledge (`--sref` numeric codes, Style Explorer, blending, `--sref random`)- `/discover-styles` command for browsing and cataloging MJ style codes- Category normalization (20 → 11 clean categories)- Pattern promotion pipeline (4 patterns graduated low → medium based on evidence)## Git State- 5 commits on main, pushed to remote- Untracked: reference-images/, sessions/, mydatabase.db## Known Issues- Context overload when reading full system (~75K+ tokens with startup hooks)- Startup hook injects ~28K tokens of observation history per conversation- AGENTS.md (56KB) + all rules = redundant double-load risk