Lohnt sich dein eigenes Reel Farm?

Build a local command-line tool that mass-produces vertical faceless short-form videos. No web UI, no accounts, no telemetry, no cloud storage.

Stack: TypeScript on Node 20, run with tsx. Use commander for the CLI, zod for config validation, and fluent-ffmpeg shelling out to a locally installed ffmpeg. SQLite via better-sqlite3 for the job queue and render history. No frontend framework, no Docker.

Secrets in .env, loaded with dotenv: OPENAI_API_KEY, ELEVENLABS_API_KEY, PEXELS_API_KEY. Fail loudly on startup if a key needed by the chosen command is missing.

Commands:
1. reelfarm script --topic "..." --count 5 : call an LLM to produce N short scripts as JSON, each with a hook line, 3 to 5 body beats, and a closing line, target 25 to 40 seconds when spoken. Store each as a pending job row.
2. reelfarm voice --job-id X : send the script to ElevenLabs TTS, save MP3 to media/voice, and normalize loudness to -14 LUFS with the ffmpeg loudnorm filter in two passes.
3. reelfarm captions --job-id X : transcribe the rendered voiceover with whisper.cpp if a local binary is configured, otherwise the OpenAI transcription API with word timestamps, and write an ASS subtitle file with 2 to 4 words per screen, centered in the middle third, thick outline, no more than 90 percent of frame width.
4. reelfarm clips --job-id X : pull vertical stock clips from Pexels matched to keywords extracted from each beat, cache downloads in media/clips, and never reuse the same clip twice within one batch.
5. reelfarm render --job-id X : with ffmpeg, scale and center-crop clips to 1080x1920, cut between them on beat boundaries, overlay the ASS captions, mix voiceover over an optional background track at -22 dB, and write out/{job-id}.mp4 with faststart, H.264, yuv420p, 30 fps.
6. reelfarm batch --topic "..." --count 10 : run the whole chain end to end, serially, resumable, skipping steps whose output already exists on disk.

Explicitly out of scope: posting or scheduling to any social platform, analytics, AI video generation models, multi-user anything. Print the output path and let the human upload.

Write a config.example.json with caption style, voice id, target duration, and clip source. Include a README with the ffmpeg install line, the .env template, and one working batch example. Add a smoke test that renders a 5 second clip from a bundled sample audio file and a solid-color background so the pipeline can be verified without spending API credits.