| name | brand-video |
| description | Produce one or many consistent, on-brand faceless marketing videos end-to-end from a topic — script → ElevenLabs voiceover (with timestamps) → one styled illustration per beat via the grounded pipeline (generateImage(), auto-detect + auto-LoRA) → ffmpeg stitch to 1080p, optionally uploaded to YouTube as Unlisted. The visual LOOK is a selectable style from styles.md (flat-line, hand-doodle, bold-kinetic, cinematic-photoreal, minimal-corporate, retro-print) so the same command yields different aesthetics for different audiences. Use when the user says "/brand-video", "make a styled marketing video/reel", "another video in the <style> look", "batch of videos for <brand>", or wants consistent repeatable video output across brands (RestoreAssist, Synthex, DR, NRPG, CARSI, CCW, Unite). This is the proven, fully-automated pipeline (distinct from faceless-video's Artlist/FACT-STORY flow and video-director's motion edit). |
brand-video — consistent styled faceless videos, on demand
Invoke as /brand-video. One command, many looks. Built from the validated
RestoreAssist run (9 videos shipped). Picks a style from styles.md, holds
everything else constant so output is consistent batch-to-batch and brand-to-brand.
Inputs (ask only if missing; infer sensibly)
- brand — RestoreAssist / Synthex / DR / NRPG / CARSI / CCW / Unite, or a client. Sets claims + CTA + voice.
- style — a key from
styles.md (default flat-line). This is the "look" / dropdown value.
- topic(s) — one angle, or a list for a batch (one video each, parallel).
- voice —
ELEVENLABS_VOICE_ID (per-brand; from ~/.hermes/.env). CRITICAL: confirm the current value before generating — the user changes it.
- upload? — default: render + stage only. Upload to YouTube Unlisted only when asked.
Pipeline (per video — one slug folder each, never share a dir)
Visual generation (binding): all images/video route through the grounded
pipeline — see .claude/rules/real-images-only.md + the grounded-visuals
skill. Direct provider calls fail CI.
- Script →
<slug>/script.md: one clean VO paragraph, ~120–140 words, ~12–16 short sentences (one sentence = one visual beat = one image), problem→solution→CTA. Ground claims in the brand wiki; never invent features. End with the brand CTA line.
- Voiceover → run the bundled tts.py (reads
ELEVENLABS_API_KEY + ELEVENLABS_VOICE_ID + DIR):
cd <workdir> && set -a && . ~/.hermes/.env && set +a && DIR=$PWD/<slug> python3 <skill>/pipeline/tts.py
Writes <slug>/voiceover.mp3 + <slug>/transcript.json (beats). Note the beat count.
- Images → one per beat via the grounded pipeline —
generateImage()/generateBatch() in lib/services/ai/image-generation.ts (or the generate_image MCP studio tool). Each prompt = the chosen style's POSITIVE token + a concrete scene for that sentence, + the style's NEGATIVE token; the industry auto-detects from the prompt and the trained LoRA (e.g. carpet-style-v1, trigger ccwcarpet) auto-applies where relevant. Landscape 16:9. Save zero-padded <slug>/images/01.png …. No owned references for the subject ⇒ the call returns blocked: true — grow the reference library (real photos), never fall back to a direct provider. Use slug-prefixed filenames to avoid parallel collisions (see [[parallel-agents-shared-output-collision]]).
- Stitch →
python3 <skill>/pipeline/stitch.py <slug> → <slug>/final-1080p.mp4 (1920×1080@25, image timing from beats). Verify final duration ≈ voiceover duration.
Batches & new looks
- Batch: dispatch one sub-agent per topic (parallel), each owning its slug dir. See the RestoreAssist run for the exact agent prompt shape.
- Re-voice (voice changed, same script/images): regenerate VO, rescale
list.txt durations by new_total/old_total, re-stitch. (revoice pattern.)
- Add a look: append a block to
styles.md (positive/negative token). That key immediately becomes a selectable style — this registry IS the dropdown's data source.
Upload (YouTube, optional)
Browser file-upload is dead (host paths rejected). Use the YouTube Data API path documented in [[youtube-api-upload-restoreassist]]: Desktop OAuth client → token at ~/.hermes/yt-restoreassist-token.json → videos().insert privacyStatus=unlisted, category 27, not-for-kids; guard on channel title; idempotent via uploaded.json. Requires YouTube Data API v3 enabled in the OAuth client's GCP project.
Guardrails
- Claims must be real (brand wiki / product truth). Flat-illustration + synthetic voice → YouTube AI-disclosure usually "No" (confirm at upload; flip to Yes if AI music added).
- Keep one style per batch for consistency. Keep
SKILL.md ≤200 lines per [[feedback-tight-code]].
- Reference implementation + assets:
/Users/phill-mac/restoreassist-videos/.
