midjourney-cc-skill: Command for Claude Code

Log Iteration

Log a Midjourney generation attempt. This captures the prompt, parameters, result assessment, and user feedback for later pattern extraction.

Instructions

1. Verify database access. Run SELECT COUNT(*) FROM sessions via sqlite-simple MCP. If the query fails, tell the user: "Database not available. Run claude mcp add sqlite-simple -- npx @anthropic-ai/sqlite-simple-mcp mydatabase.db then restart Claude Code." Do not proceed without database access.

2. Check for an active session. Query the database:

   SELECT id, intent, total_iterations FROM sessions WHERE status = 'active' ORDER BY created_at DESC LIMIT 1
   

3. If no active session exists, ask the user:

   • What are you trying to create? (intent)
   • Do you have a reference image to describe? (reference_description)
   • Then create a new session with a generated UUID.

4. Determine what data is available. There are five scenarios:

   Scenario A: Browser automation capture (richest data path). If the generation was submitted via Playwright MCP (from /new-session or manual browser control):

   • Create the iteration screenshot directory:

     mkdir -p sessions/{session_id_first_8}/iter-{NN}/
     

   • Use batch capture via browser_run_code to capture all 4 images in one tool call (see "Batch Image Capture" in rules/auto-core-workflows.md). This saves ~40-50% context compared to individual navigate/screenshot calls:

     browser_run_code({
       code: `async (page) => {
         const jobId = '[JOB_ID]';
         const dir = 'sessions/[ID]/iter-[NN]';
         for (let i = 0; i < 4; i++) {
           await page.goto('https://www.midjourney.com/jobs/' + jobId + '?index=' + i);
           await page.waitForTimeout(3000);
           await page.screenshot({ path: dir + '/img-' + (i + 1) + '.png' });
         }
         return 'All 4 images captured';
       }`
     })
     

   • You already know the exact prompt and parameters (from the session)
   • Analyze all 4 images in the grid using the 7 standard scoring dimensions (subject, lighting, color, mood, composition, material, spatial) — see rules/core-assessment-scoring.md. Score every dimension for every image, even if "not applicable" (score 1.0).
     • Flag any dimensions where agent confidence is low (especially spatial_relationships)
     • Identify the best candidate and explain why
     • Note consistency patterns: if all 4 miss the same thing, it's a prompt-level issue; if they diverge, it's MJ interpretation variance
   • Formulate gap analysis per image and overall
   • Present scores as PRELIMINARY — ask the user to validate before logging, especially for low-confidence dimensions
   • After user confirms/corrects, recommend next action using the Iteration Action Decision Framework: Upscale best, Vary promising (specify which image), or prompt edit
   • If user approves an action, perform it via browser and capture the result