tiktok-product-video-machine — Next.js SaaS Template

SHA-256: 3d827c55f3f42367840c687f819ccefd5672c8fbe7d7ae5322b2e1f4c9272181

2026-10-02T01:01:57.900Z

# TikTok Product Video Machine

Watches TikTok Creative Center + Kalodata for verified-seller winning products and their new brand-family SKUs, generates original AI videos with Higgsfield, and auto-posts to your accounts on an adaptive cadence.

Sibling to `booktok-trailers/` in this repo — same architecture, different domain.

## Quickstart (30–45 minutes of clicks)

```bash
pnpm install
pnpm exec playwright install chromium
cp .env.local.example .env.local
# Fill in Supabase, Redis, Stripe, Higgsfield, AWS keys.

# 1) Grab your Kalodata free-tier session cookie (opens a browser)
pnpm grab-kalodata-cookie

# 2) DB
psql "$SUPABASE_DB_URL" -f supabase/migrations/0001_init.sql

# 3) Seed one TikTok account with encrypted cookies
pnpm seed-cookies dkgrissomshop bookish

# 4) Start dev + worker in two terminals
pnpm dev            # terminal 1: Next.js
pnpm worker:dev     # terminal 2: BullMQ worker

# 5) Full smoke test
pnpm smoke
```

## What the machine does, per day

1. **Scans** Creative Center for the top verified-seller products (worker/scan.ts).
2. **Expands** each hit into its brand family via Kalodata to catch new SKUs from proven shops.
3. **Generates** 3–8 AI videos per product across camera presets (worker/generate.ts → lib/higgsfield.ts).
4. **Captions** each with a hooked hashtag stack, run through moderation deny-list (worker/caption.ts).
5. **Posts** on adaptive cadence — Playwright today, official API when approved (worker/post.ts → lib/posting/*).
6. **Rolls up** 1h and 24h reach ratios, adjusts each account's daily cap (worker/metrics.ts).

## Compliance guardrails

- Never re-uploads another creator's video — extracts product images and generates fresh.
- AI-generated tag rendered into every output (add to Higgsfield prompt or via ffmpeg overlay).
- Rate limits per account: 10/day ceiling, 45-min spacing, 14-day warmup.
- Verified-seller filter on by default.
- Session cookies encrypted with `pgp_sym_encrypt` server-side.

## Repo shape

```
tiktok-product-machine/
├── app/                     # Next.js 15, App Router
│   ├── api/{scan,generate,videos,post,webhook,stripe}/route.ts
│   ├── dashboard, queue, accounts, case-studies/
│   └── page.tsx (landing + pricing)
├── lib/
│   ├── signal/              # creative-center + kalodata + brand-family
│   ├── extraction/          # product images + seller badge normalization
│   ├── prompts/             # product-video prompt + caption LLM
│   ├── posting/             # cadence + official + unofficial + router
│   ├── moderation/          # deny-list
│   ├── metrics/             # reach-ratio rollup
│   ├── higgsfield.ts        # transport
│   └── comfyui.ts           # Phase-2 stub
├── worker/                  # BullMQ jobs
├── supabase/migrations/     # RLS-on schema
├── scripts/                 # seed-cookies, warmup, smoke-test
└── tools/                   # grab-kalodata-cookie (interactive)
```

## Path C phases

- **Phase 1** — Higgsfield API only. Validate. Eligible for the $50K bounty.
- **Phase 2** — Flip `COMFYUI_LOCAL_FRACTION=0.5`. Half the jobs go to your GPU box running Wan 2.2 + FLUX Schnell. Cost per video drops ~70%.
- **Phase 3** — `COMFYUI_LOCAL_FRACTION=1.0`. Higgsfield becomes failover.

Details in `../../build-specs/05-tiktok-product-machine.claude.md`.