anabologyco-maker/claudia-bot/plates

Plates - lead, book, Grok runs Midjourney, ingest, picks, fixes

Midjourney (Studio, Director): read docs/50-plates-midjourney.md §1 first. Grok runs the prompt blocks in its own browser and downloads every image (slop relay job --kind midjourney, §9); the owner only judges the picks.

Lead first (G3)

1. Write the canon with the owner (§5.1, §5.3): framing first, visible garments, ethnicity, age, skin; plus a short far version for wide shots. slop --json -p NAME lead --canon '<canon>' --far '<far version>'.
2. Midjourney's switches are the owner's, asked once per project (defaults: Raw on, personalization on, HD off): slop --json -p NAME project set mj.raw true|false (likewise mj.personalize, mj.hd). Never a --v or a profile code in a prompt: Midjourney's default is its newest model, and Grok presses the P button.
3. 4 test plates of the lead (CU, MCU, FS, WS). Before a board exists: write the 4 prompts yourself (§4.9 template, the canon verbatim, the far version for FS and WS) and ask Grok for them: slop -p NAME relay job --kind midjourney --text '<prompt>' --plate lead-cu (a --text and a --plate per plate). Ingest each download: slop --json -p NAME plates ingest FILE --plate lead-cu (then lead-mcu, lead-fs, lead-ws), then slop --json -p NAME plates sheet --plates lead-cu,lead-mcu,lead-fs,lead-ws. Once a board exists use slop --json -p NAME plates book --batch NAME-00-lead --only K1,K2,K3,K4 instead (slop --json -p NAME plates status lists the keys). "Is this the lead? Flag any wrong face." Record: slop --json -p NAME decide '<their words>' --by owner.

The book (§7, §8)

1. Settings: the version stays Midjourney's default (its newest); Raw and HD are in the params, personalization is the P button (the relay job applies the owner's choices).
2. slop --json -p NAME plates book --batch NAME-01-core --plan standard (their plan: basic, standard, pro or mega). Read every prompt against the lint list (§7.3). Grok runs one Midjourney job per line (never a permutation block), so a book needs no paste test.
3. Show the owner plates/prompts.md (G5 when it spends money: an API batch). Then one relay job for the Midjourney lines: a --text per line of plates/mj/BATCH.lines.txt and its --plate (the keys in plates/mj/BATCH.json).

Ingest and pick (§10)

1. Grok's downloads arrive in inbox/midjourney/ named PLATE-N.png: slop --json -p NAME plates ingest inbox/midjourney --batch BATCH. Unmatched images: send plates/unmatched.jpg, the owner confirms, run it again with --map KEY=PLATE.
2. slop --json -p NAME plates sheet, send the sheets, the owner replies like "hall-ms #2, hall-wide #4": slop --json -p NAME plates pick 'hall-ms#2, hall-wide#4'. "You pick": slop --json -p NAME plates pick --all-first, and they review on the board. Then check each sung shot's pick: the whole mouth, nothing in front of it (no microphone, hand or hair); else pick another.
3. slop --json -p NAME plates vary (one image per shot; no two clips from one start frame). Gaps: slop --json -p NAME plates status.

Fix the source, never the frame (§12)

Wrong face or framing: re-prompt and re-roll. One wrong detail on a good pick: an image edit (about $0.14, the owner's yes first): cd /workspace/slop/projects/NAME && slop --json -p NAME gen edit plates/cands/PLATE/2.jpg plates/cands/PLATE/5.png --change '<only this, target state>' --keep '<what must not move>' --yes, then slop --json -p NAME plates pick PLATE 5. Check the canon again after every edit. A declined image is never edited (docs/62-moderation.md §1).

API stills (Quick tier, or no Midjourney; docs/51-plates-api.md §2.5, §3)

slop --json -p NAME plates book, then slop --json -p NAME gen image --jobs (a dry run with the estimate; the lead's 4 test plates first with --only K1,K2,K3,K4, G3). The owner reads the book and the estimate; on yes slop --json -p NAME gate prompts --approve --note '<their words>' and slop --json -p NAME gen image --jobs --yes. Identity anchor, after G3: slop --json -p NAME plates pick PLATE K (the approved close-up), then slop --json -p NAME plates book --anchor PLATE (the full book plates/prompts.md, which G5 needs) gives every lead plate's job it as reference; gen image --jobs --missing skips plates already made (one image: gen image with --ref PLATE:K). The owner re-picks the close-up: that plates book --anchor PLATE again; text canon only again: slop --json -p NAME plates book --no-anchor. Images from Higgsfield: slop --json -p NAME plates ingest --url 'URL' --plate PLATE; from GenerateImage (no key): saved as uploads/gen/PLATE/K.png, then slop --json -p NAME plates ingest uploads/gen (both: skill mv-quick step 7). Named regions for devices: slop --json -p NAME plates box PLATE --label 'a face' --uv 0.41,0.18,0.16,0.30 (§6.4).