project creatortrack ai voice profile pipeline 20260715T172503

---

name: CreatorTrack AI voice-profile / script-generation pipeline for Gus brand description: Multi-week build: ingest video history across YouTube/Instagram/Facebook/TikTok, extract per-video structure, cross-post match, then train a voice profile to generate and eval new scripts. type: project

---

CreatorTrack is building an AI content-generation pipeline for the Gus Outdoor Co brand (bass channel, @gus_the_bass / Gus Junior) on branch feat/social-meta-ingest (Meta lane) and feat/social-data-layer-and-scripts (main lane), worked in parallel warm-slot worktrees.

Pipeline stages: (1) per-platform video ingestion/backfill (YouTube 548, Instagram 505, Facebook 699, TikTok 887 = 2,639 total videos for ws 85), (2) transcript ingestion + per-video structure extraction via Claude Haiku (hook/beats/cta/tone/signature phrases into raw.structure), (3) cross-post matching (app.social_videos.crosspost_of_id) to link the same video across platforms via title/duration/date-proximity scoring, deferring an embedding tier, (4) a voice-profile synthesis step (Task 14) trained on the full structured/matched corpus, then (5) generation + holdout eval (Tasks 15-18).

Why: this is the first AI-native content tool built on top of CreatorTrack's social data layer, aimed at scaling Gus's per-video scripting workflow using his own historical voice/style.

How to apply: when picking this back up, check comms/results/creatortrack-social-*-m*.md and comms/results/creatortrack-social-data-layer-and-scripts-app-task-*.md result files, and the feat/social-meta-ingest / feat/social-data-layer-and-scripts branches, before assuming the pipeline's state from memory alone; it ships as two branches (rebase order matters) via the standard CT ship sequence.

[auto-memory session 60d132de-4659-485c-8044-427af0ccce7f, confidence 0.70, mode staged]