anabologyco-maker/claudia-bot/mv-quick

Quick tier - song to finished video in one sitting

Read docs/00-overview.md §2.1 and §3 once. Loop: slop --json -p NAME status, then the step its next.cmd names. Free automatic steps: slop --json -p NAME next --steps 3; a job: the profile's turn rule.

One owner message each: lyrics + lead + look; board sheet (G4) + the lead stills' estimate and the total to v1; lead sheet (G3) + the rest of the book and its estimate (G5); picks ("I'll pick unless you want to"); clip book + estimate; the excerpt; v1 + 9:16 choice + posting copy. "Go on to v1" at the board: record it with decide, then run every step to v1 by the turn rule within that total (you pick the lead close-up and name it); stop only at a failure or a cost over it. After each gate, offer a new chat for the next stage.

No key at all (no OpenRouter, no Higgsfield): their own song file, stills from Grok's GenerateImage (step 7), no clips (2.5D and devices only: slop --json -p NAME project set budget.motion_s 0 before the board), no "say so" checks.

1. Song (docs/10-song.md §6, §9).

• Upload: skill song step 7. Ask for the lyrics in the same message (Suno: the Lyrics field). An upload needs no G2.
• No song, has OpenRouter: first one message: what it is about, genre, singer, with defaults (§1 questions 1, 3, 4). Write the lyric by skill song step 3, then Lyria by its Lyria section. Send audio/lyria-N.mp3 for G2 (a retake: $0.08, budget 3-5).

1. Lyrics as sung (numbers spelled out) into lyrics/lyrics.txt (docs/20-analysis.md §1); slop --json -p NAME lyrics lint.
2. Analysis needs the full install: bash /workspace/claudia/bin/setup.sh --check --full --json ok (not yet: steps 2 and 5 meanwhile, then check again). slop --json -p NAME job start -- analyze (2-5 min), then skill song step 8 (ear check, bars, sections).

• Lyria: slop --json -p NAME lyrics check --yes (about $0.002; say so). Show the changed lines; on yes slop --json -p NAME lyrics check --apply-lines N,M, then slop --json -p NAME job start -- analyze --steps align.

1. Display text: skill song step 9.
2. Lead and look, one message (docs/70-motion-design.md §1). Lead: slop --json -p NAME lead --canon '<identity text>' --far '<short far version>', or slop --json -p NAME lead --none plus slop --json -p NAME project set board.lead_min 0. Video on OpenRouter only: offer a stylized lead (docs/51-plates-api.md §3.1). Look, with defaults: generator colours (native) / one clean grade / print; density SP-budgeted / dense / sparse; one accent colour not in the footage's main hue. Record with decide, then slop --json -p NAME project set look.accent '#HEX' and slop --json -p NAME project set look.mode MODE (native, clean-grade or print). No grade or mono words in still prompts.
3. Board: skill storyboard at the quick pace (docs/40-storyboard.md §6.1). G4 yes: slop --json -p NAME board approve --note '<their words>'.
4. Stills (docs/51-plates-api.md §2.4, §2.5, §3), lead first:

• Production model before plates book: slop --json gen models --kind image; the §2.4 final model listed: slop --json -p NAME project set models.image openai/gpt-image-2.5-flare. Stills on their ChatGPT plan ($0): skill setup-agents, then every gen image below without --quality medium.
• Lead: 4 plates whose shots show the lead in CU, MCU, FS, WS (cd /workspace/slop/projects/NAME && jq -r '.shots[] | "\(.id) \(.framing) \(.plate) \(.lead_on_screen)"' board/shots.json): slop --json -p NAME plates book --only K1,K2,K3,K4, slop --json -p NAME gen image --jobs --only K1,K2,K3,K4 --quality medium (dry run: the estimate), owner yes, the same with --yes, then slop --json -p NAME plates sheet. G3: "Is this the lead? Flag any wrong face." With it the rest (G5): the full book (slop --json -p NAME plates book) and the estimate of slop --json -p NAME gen image --jobs --missing --quality medium.
• On yes: slop --json -p NAME plates pick PLATE K (the close-up they approve), slop --json -p NAME plates book --anchor PLATE (lead plates anchored on it), slop --json -p NAME gate prompts --approve --note '<their words>', then that gen image with --yes.
• Higgsfield, no OpenRouter: price one plate (generate_image with get_cost), the book and the total for the G5 yes; then generate_image_batch (up to 12 a call; anchored jobs with the close-up as reference, uploaded as in skill clips step 8), jobs_wait {timeout_seconds: 15} until done, slop --json -p NAME plates ingest --url 'URL' --plate PLATE per result; check transactions (the ledger misses them).
• No key, GenerateImage ($0) in place of gen image: per plate of the step without candidates, its prompt (16:9) and refs (as references) from plates/jobs.json; save each image as /workspace/slop/projects/NAME/uploads/gen/PLATE/K.png (K = 1, 2...), then slop --json -p NAME plates ingest uploads/gen.
• Picks: slop --json -p NAME plates sheet, the owner picks (or "you pick": slop --json -p NAME plates pick --all-first), then slop --json -p NAME plates vary.

1. Clips (optional, 2-4; skill clips): run the audit's clip_cmds lines. Motion: add --model bytedance/seedance-2.0-mini --res 480p. Sung, only if the owner wants singing on screen: illustrated lead --model bytedance/seedance-2.5 --res 480p (about $0.52 per 5 s); photoreal lead or no OpenRouter --provider higgsfield --res 480p (3 credits a second). Never a sung clip on 2.0 Mini or Wan (no lip copy). Else skip sung lines: the plates carry them in 2.5D.
2. Render and deliver: skill render-and-review.

Promote later: slop --json -p NAME project set tier studio (needs OpenRouter or Higgsfield); the Studio caps only after the owner's yes: slop --json -p NAME project set budget.openrouter_usd 100 --owner-ok --by owner (budget.higgsfield_credits 1500 alike).