VideoGenAI

An end-to-end pipeline that turns a one-line topic into a researched, scripted, animated, voiced YouTube video — with a human-approval cockpit between every stage.

Multi-channel by design. Channel behavior (tone, source-balancing rules, jargon depth, visual style) is config-driven, not hardcoded. Adding a new channel is a config file, not a code change.

Status

Stages 1–8 complete and wired. Full text pipeline (brief → research → jargon → script → fact-check → storyboard) plus asset generation (TTS + stock video + AI stills) and render manifest assembly are production-ready. QA and publish (stages 9–10) are next.

Phase	Status
DB schema + migrations	Done
LLM adapter (Anthropic + OpenAI kill switch)	Done
Brief builder	Done
Researcher (Brave Search + page fetch)	Done
Jargon miner	Done
Scriptwriter	Done
Fact checker	Done
Storyboarder	Done
Inngest orchestration + approval gates	Done
Cockpit UI (run list, detail, approvals)	Done
Cockpit revision flow (edit JSON, re-run w/feedback)	Done
Pipeline pause / resume kill switch	Done
Asset generation (ElevenLabs + Pexels + Flux)	Done
Assembler (render manifest + asset copy)	Done
Cockpit asset preview (audio, image/video grid)	Done
Retry assets button (re-run stages 7–8 only)	Done
QA stage	Next
YouTube publish	Next

Pipeline

flowchart TD
    USER([Operator]) -->|topic + channel| NEW_RUN

    subgraph COCKPIT["Cockpit (Next.js 15)"]
        NEW_RUN[New Run Form]
        RUN_LIST[Run List]
        RUN_DETAIL[Run Detail]
        APPROVE[Approval Queue]
    end

    subgraph INNGEST["Inngest — Durable Orchestration"]
        NEW_RUN -->|POST /api/runs| API_CREATE[Create Run API]
        API_CREATE -->|videogenai/run.start| ORCH

        subgraph ORCH["pipelineRun function"]
            direction TB
            S1[1 · Brief Builder]
            S2[2 · Researcher]
            S3[3 · Jargon Miner]
            S4[4 · Scriptwriter]
            S5[5 · Fact Checker]
            S6[6 · Storyboarder]
            S7[7 · Assets ✅]
            S8[8 · Render ✅]
            S9[9 · QA ⬜]
            S10[10 · Publish ⬜]

            S1 -->|awaits approval| G1{Gate}
            G1 --> S2
            S2 -->|awaits approval| G2{Gate}
            G2 --> S3
            S3 -->|auto-approved| S4
            S4 -->|awaits approval| G4{Gate}
            G4 --> S5
            S5 -->|awaits approval| G5{Gate}
            G5 --> S6
            S6 -->|awaits approval| G6{Gate}
            G6 --> S7
            S7 --> S8
            S8 --> S9
            S9 -->|awaits approval| G9{Gate}
            G9 --> S10
        end

        APPROVE -->|POST /api/runs/:id/stages/:stage/approve| APPROVE_API[Approve API]
        APPROVE_API -->|videogenai/stage.approved| ORCH
    end

    subgraph DB["Neon Postgres (Drizzle ORM)"]
        RUNS[(runs)]
        STAGES[(stages)]
        ASSETS[(assets)]
    end

    subgraph LLM["LLM Layer"]
        SWITCH{Provider?}
        ANTHROPIC[claude-opus-4-7]
        OPENAI[gpt-4o / codex]
        SWITCH --> ANTHROPIC
        SWITCH --> OPENAI
    end

    subgraph EXTERNAL["External APIs"]
        BRAVE[Brave Search]
        ELEVENLABS[ElevenLabs TTS]
        PEXELS[Pexels Stock Video]
        FLUX[Flux via Replicate]
        YOUTUBE[YouTube Data API v3]
    end

    ORCH -->|reads/writes| DB
    ORCH -->|generateStructuredOutput| LLM
    S2 --> BRAVE
    S7 --> ELEVENLABS
    S7 --> PEXELS
    S7 --> FLUX
    S10 --> YOUTUBE

    RUN_DETAIL -->|GET /api/runs/:id| ORCH
    RUN_LIST -->|GET /api/runs| DB

Loading

Stage details

#	Stage	LLM	Approval	Output
1	Brief Builder	Yes	Human	Angle, audience note, must-cover/avoid, tone
2	Researcher	Yes	Human	FactPack: sources + verified facts
3	Jargon Miner	Yes	Auto	Term → definition pairs
4	Scriptwriter	Yes	Human	Scenes with text + citations + duration
5	Fact Checker	Yes	Human	Score, verdict, issues list
6	Storyboarder	Yes	Human	Scenes with visual briefs + shot types
7	Assets	—	—	narration.mp3 + b-roll clips + AI stills manifest
8	Render	—	—	render-manifest.json + asset directory (MP4 via CLI)
9	QA	Yes	Human	Quality report + issues
10	Publish	—	—	YouTube upload + metadata

Cockpit features

• Run list — all runs with status chips; click to open detail
• Stage cards — collapsible, show output, metrics, timestamps
• Approval gates — approve as-is, edit JSON inline, or re-run stage with feedback
• Asset preview — narration audio player, 16:9 image/video grid with hover-play
• Render manifest viewer — resolution, duration, asset paths
• Retry assets — re-run stages 7–8 without touching LLM stages (saves API cost when asset keys are newly added)
• Pause / resume — pause the pipeline at any inter-stage boundary; resume picks up from the exact same point (Inngest waitForEvent, zero compute while paused)
• Pipeline visualiser — glossy stage nodes with per-state color glow (done=green, running=cyan, approval=amber, failed=red)
• LLM provider toggle — switch between OpenAI and Anthropic in the header without restarting

Stack

• Orchestration: Inngest (durable steps, human-approval gates, pause/resume, 14-day timeouts)
• LLM: Anthropic Claude / OpenAI GPT — switchable via cockpit toggle or LLM_PROVIDER env
• Video: Remotion (React-based video composition — run locally via CLI)
• Cockpit: Next.js 15 App Router
• DB: Neon Postgres + Drizzle ORM
• Voice: ElevenLabs (free-tier Adam voice; bring your own voice ID on paid plan)
• Visuals: Pexels stock video + Flux Schnell via Replicate
• Search: Brave Search API
• Publish: YouTube Data API v3

Quickstart

Requires Node 20+ and pnpm 9+.

pnpm install
cp .env.example .env
pnpm --filter @videogenai/db db:migrate   # apply schema to Neon
pnpm dev                                  # cockpit on :3000, Inngest dev on :8288

Then open localhost:3000/runs/new, pick a channel, enter a topic, and watch the pipeline run.

Required API keys

Key	Used by	Get it at
DATABASE_URL	DB (all stages)	neon.tech
OPENAI_API_KEY or ANTHROPIC_API_KEY	LLM stages 1–6	platform.openai.com / console.anthropic.com
BRAVE_SEARCH_API_KEY	Stage 2 researcher	api.search.brave.com
ELEVENLABS_API_KEY	Stage 7 TTS	elevenlabs.io
PEXELS_API_KEY	Stage 7 stock video	pexels.com/api
REPLICATE_API_TOKEN	Stage 7 AI stills (Flux Schnell)	replicate.com

ELEVENLABS_DEFAULT_VOICE_ID is optional — defaults to Adam (pNInz6obpgDQGcFmaJgB), which works on the free tier. Library voices require a paid ElevenLabs plan.

To render the assembled video as MP4, run Remotion locally after stage 8 completes:

npx remotion render apps/web/public/assets/runs/<run-id>/render-manifest.json

Claude Code skills

Custom slash commands for working on this repo — run them in Claude Code:

Command	What it does
/add-stage	Scaffolds a new pipeline agent stage (Zod schema, LLM call, db-ops, approval gate)
/add-channel	Creates a new channel config following the existing schema
/db-migrate	Applies pending Drizzle migrations to Neon
/run-status	Shows recent pipeline runs and any stages awaiting approval

Monorepo layout

apps/
  web/              # Next.js 15 cockpit (UI + API routes + Inngest handler)
packages/
  db/               # Drizzle schema, migrations, Neon client
  pipeline/         # Agent stages, Inngest orchestration, LLM adapter, skills
  types/            # Shared Zod schemas and TypeScript types
docs/
  PLAN.md           # Phased build plan
  ARCHITECTURE.md   # Stage details and data flow
  CHANNELS.md       # Channel config schema
.claude/
  commands/         # Claude Code custom slash commands

Channels

1. Aussie politics explained for dummies — explainer style, jargon-defined, source-balanced, no political skew
2. Latest tech explainers — fresh news ingested, technical terms unpacked

More channels added as YAML configs once the pipeline stabilizes.

Docs

• PLAN.md — phased build plan
• ARCHITECTURE.md — stage details, agents, data flow
• CHANNELS.md — channel config schema
• AGENTS.md — guide for AI coding assistants on this repo

License

Proprietary — all rights reserved.