========================================================================
 AI VIDEO PIPELINE — PROJECT NOTES & RESUME GUIDE
 (Kids YouTube channel — image-to-video production)
 Last updated: 2026-07-17
========================================================================

------------------------------------------------------------------------
HOW TO RESUME THIS WITH CLAUDE
------------------------------------------------------------------------
1. Start Claude Code from THIS folder:  /home/working2/public_html
   (Memory is tied to this directory — it auto-loads here.)

2. Use one of these prompts:
     "Continue the AI video pipeline project for my kids YouTube channel."
     "Let's continue the Wan 2.2 video pipeline discussion."

3. If Claude seems not to remember, say:
     "Check your memory for the AI video pipeline project."

------------------------------------------------------------------------
THE GOAL
------------------------------------------------------------------------
- Kids' YouTube channel (Numberblocks / PJ Masks / Paw Patrol / Bluey
  "reimagined as X" style). ~10-minute videos, target ~10 videos/month.
- Current workflow:
     ChatGPT (make images) -> Canva AI (image-to-video, many 5s clips)
     -> CapCut (stitch into the long video).
- Problem: Canva AI video is throttled (~4-min hold between clips) AND
  monthly-credit-capped -> can't scale to full videos cheaply.
- Sample clips already in this folder:
     "PJ Masks Take Care of Baby Animals 1.mp4"
     "PJ Masks Take Care of Baby Animals 2.mp4"
     (4K / 5s / silent AI image-to-video with real motion.)

------------------------------------------------------------------------
THE RECOMMENDATION WE LANDED ON
------------------------------------------------------------------------
- MOTION step: rented HOURLY RTX 4090 + Wan 2.2 (TI2V-5B) + Lightning
  LoRA, output 1080p (SKIP 4K upscaling — YouTube doesn't need 4K).
     Cost: ~$18-40 / month for 10 videos.
     Flat monthly (~$159 RTX PRO 4000) only worth it above ~25 videos/mo.
- VOICE: Edge-TTS (free, unlimited, good Hindi/English).
- STITCH / edit: CapCut or ffmpeg (free).
- BIGGEST cost lever = clip count. Reuse clips / hold frames ->
  ~60-80 unique clips per video, NOT 120. This ~halves cost.
- Free Kaggle GPU (~360 clips/mo) only covers ~1/3-1/2 of the load ->
  optional overflow, not the backbone.

------------------------------------------------------------------------
STRATEGY / COPYRIGHT NOTE
------------------------------------------------------------------------
- Current videos use trademarked characters -> real takedown +
  demonetization risk under YouTube's 2025/26 "inauthentic / mass-
  produced content" policy.
- Advised pivot: ORIGINAL characters (working name "Countlings") +
  real narration, keeping the proven "reimagined as X" format.

------------------------------------------------------------------------
OPEN NEXT STEPS
------------------------------------------------------------------------
1. Count the actual unique clips in ONE real 10-min video -> get a
   precise monthly cost (instead of a range).
2. Pick GPU provider: RunPod (easiest) vs Vast.ai (cheapest).
3. Decide whether Claude starts BUILDING the pipeline scripts.
   NOTE: "don't change coding" = don't touch the freejobalert site.
   Building the pipeline = NEW scripts in a separate folder; still
   needs an explicit "yes, build it" from you.

------------------------------------------------------------------------
ENVIRONMENT FACT
------------------------------------------------------------------------
- This server CANNOT generate video (no GPU, Python 3.6). It can only
  do ffmpeg assembly/stitching. Generation must use an external GPU/API.
========================================================================