anabologyco-maker/claudia-bot/song

Song - concept, lyrics, Suno pack, take, analysis

Read docs/10-song.md §0 first. Suno has no API: Grok creates each take in its own browser (slop relay job --kind suno, docs/10-song.md §5); the owner only listens and picks.

1. Questions (§1): one message with the 9 questions and their defaults. Record the answers verbatim: slop --json -p NAME decide '<answers>' --by owner. Words they never want: cd /workspace/slop/projects/NAME && jq '.banned=["WORD1","WORD2"]' project.json > p.tmp && mv p.tmp project.json.
2. Concept (§2): web search for a fact bank (§2.1), then 3 pitches of at most 120 words: premise, through-line device, lead, arc, one risk (§2.3). G1: "Which one? (Default: A.) Anything to change?" Record the answer with slop --json -p NAME decide '<their words>' --by owner.
3. Lyrics (§3): pick a structure template (§3.2), write lyrics/lyrics.txt spelled as sung (§3.4), section tags on their own lines. slop --json -p NAME lyrics lint (add --ban WORD per banned word), fix every flagged line, lint again until it is clean (§3.7).
4. Suno pack (§4): write suno/style.txt, suno/exclude.txt, suno/lyrics-field.txt, suno/title.txt and suno/settings.md. Grok makes it in its browser: slop -p NAME relay job --kind suno --title '<title.txt>' --style '<style.txt>' --exclude '<exclude.txt>' --lyrics-file suno/lyrics-field.txt --note '<settings.md in one line>'; both takes come back in inbox/suno/.
5. Rounds (§5): the owner gives a verdict per take; change one thing; next round. Log suno/takes.md. G2: the owner names the take; record it the same way.
6. Terms (§8), once per owner: commercial use needs a download on a paid plan; the clause about not powering other AI tools is ambiguous for a video pipeline, so suggest a legal read before monetising; Free plan downloads are non-commercial (offer Lyria or an upgrade). Record their answer the same way.
7. The file (§6.2): the take the owner named, from inbox/suno/ (Grok saved both); the owner's own song: the file Grok saved from their upload, or a Drive or Dropbox share link: slop --json -p NAME fetch 'LINK' --to audio. Never take audio from Suno's CDN or by recording the player. Then slop --json -p NAME song ingest FILE (FILE = the full path; after fetch, the file it printed).
8. Analysis (docs/20-analysis.md §0; needs the full install): slop --json -p NAME job start -- analyze, then wait by the profile's turn rule. Read analysis/report.md.

• Ear check every flagged item (§3.2); fix a word: slop --json -p NAME lyrics fix --word WORD_INDEX --t0 SECONDS --note '<why>'.
• Bars (§4.3): slop --json -p NAME song bars; wrong phase: slop --json -p NAME song bars --phase PHASE.
• Sections (§5): slop --json -p NAME song sections; wrong: slop --json -p NAME song sections --set 'SECTION@START,...' built from its table (each name with its --set START value, changed where the owner says), EVERY section: it replaces all.
• Sung choruses (§2.2): copy analysis/words.json to analysis/words-w2v.json, then slop --json -p NAME job start -- analyze --steps align --whisper-fallback --yes (about $0.002; say so).

1. Display text (docs/20-analysis.md §7): slop --json -p NAME lyrics display --init (never --force: it wipes edits), edit lyrics/display.txt as it should read on screen (AGI|3, +25%|3, the accents), then slop --json -p NAME lyrics display until every line consumes its words.
2. One message to the owner: the take, tempo range, the fixes and the sections. Record their yes the same way. Then skill storyboard.

Lyria (no Suno, up to about 3 minutes, needs OpenRouter; docs/10-song.md §9)

Write suno/lyria-prompt.txt from the pack (the table in §9). Show the prompt and the price. On yes: slop --json -p NAME gen song --prompt-file suno/lyria-prompt.txt --clip --yes (30 s test, about $0.04), then slop --json -p NAME gen song --prompt-file suno/lyria-prompt.txt --as-song --yes (about $0.08; writes audio/song.wav). The owner auditions it (G2). After the analysis: slop --json -p NAME lyrics check --yes shows what the take really sings.