Voice fidelity is the moat. Read the bundled references first:
references/workspace-context.md (read-context-first + write-back),
references/account-resolver.md (auth + account_id vs row id),
references/research-context.md (live signals).
Procedure
Diarize the artist's voice into a checkable fingerprint (length/caps/emoji/
punctuation/lexicon + an explicit "avoid" list) from context/artist.md +
context/audience.md, or fall back to 10–30 real past captions
(/artists/{ROW_ID}/posts) as few-shot anchors. Draft Count (default 3) distinct
angles through the fingerprint. Verify each against the fingerprint + an
anti-slop checklist (no "Get ready", "Mark your calendars", emoji stacks, lines that
fit any artist); drop/redraft failures. Never invent a voice — no context + no posts
→ ask. Persist to content/captions/.
Guardrails
- Never fabricate the voice — diarize from real context/posts or ask.
- Stop at the text — no posting/scheduling.
References
references/workspace-context.md·references/account-resolver.md·references/research-context.md