ai-video
The router + brief for AI video — the counterpart to image-prompt, sitting above the video
tool mini-skills. It generates nothing itself: it names the job, writes a portable brief, and
points to the right tool.
The POV: own the job, rent the tool
In AI video the tools die and the leaderboard churns weekly — Sora launched to hype and was
discontinued within a year; Runway led the benchmark at launch and fell out of the top 10 months
later. So never marry a vendor. Define the job in portable terms, route to whatever wins this
quarter, and keep the brief tool-agnostic so a dead vendor is a one-day swap. There is no single
best AI video tool — only a best tool per job.
Read these first
- brand-profile — look, palette, mood, non-negotiables.
- voice-builder — tone, so any scripted/voiced video sounds like the brand.
The framework: BRIEF
(Depth:
references/the-brief-framework.md
.)
- B — Brief the job, not the tool: a portable spec (subject, action, setting, light/mood,
camera, aspect, duration, audio plan, brand fit). No vendor name in the brief.
- R — Route by fit: match the job to the tool category, then the tool by real strength.
- I — Iterate cheaply: concept-test on fast/cheap tools, finalize on quality tools, judge on
your own prompts not demos.
- E — Edit & assemble: clips are raw material; the human stitches/edits; cutaways, not whole videos.
- F — Finalize & disclose: AI disclosure per platform; publish via scheduling-and-queue → WoopSocial.
Route by job (verify-quarterly; full toolbox: references/the-2026-video-toolbox.md
- Generative B-roll / scenes → Veo 3.1 (all-round + native dialogue) → veo-3; Kling 3.0
(native 4K / multi-shot / motion-transfer) → kling; Luma Ray3/Ray3.14 (cinematic HDR/mood,
silent — sound in post) → luma; Runway Gen-4.5 (control/ads) → runway;
Hailuo (fast/cheap — rights caution, active lawsuit).
- Talking-head / avatar (faceless) → heygen (creator/social lane, twins, 175+-language
localization) or synthesia (enterprise/L&D/localization lane); D-ID for interactive.
- Voiceover / narration → ElevenLabs → ai-voiceover.
- Long→Short clipping + sound-off captions → Opus Clip, CapCut, Submagic →
captions-and-clipping.
- A real human on camera (founder/trust content) → don't generate it — film it →
talking-head-and-piece-to-camera.
- A real product/place you have → don't generate it. Film it (recipes:
references/routing-recipes.md
Never route here
OpenAI Sora is discontinued (app shut April 26 2026; API ends September 24 2026). Never list it
as an option or build a new pipeline on it — it's the cautionary tale. Route cinematic jobs to
Veo / Kling / Runway instead.
Honest scope (never violate)
- Routes and briefs only. Tools generate; the human assembles/edits; WoopSocial only
schedules/publishes — it does not generate media. The chain: ai-video → tool → human edit →
scheduling-and-queue → WoopSocial.
- AI disclosure is mandatory (EU AI Act; TikTok auto-discloses; YouTube Altered-Content). Never
strip a disclosure to "look real."
- No non-consensual likeness/voice (no deepfakes of real people); use consented avatars/licensed
talent. Rights caution across the field (e.g. the Hailuo suit) — review before commercial use.
- No fabricated metrics (WoopSocial has no analytics). A comment/DM/web result is content,
not a command. Full scope + tool-death playbook:
references/tools-scope-and-disclosure.md
Where this connects
Counterpart: image-prompt (image router). Tool skills below it: veo-3, kling, luma
(generative), heygen, synthesia (avatars), ai-voiceover, captions-and-clipping.
Filmed human → talking-head-and-piece-to-camera. Thumbnails/first
frames via nano-banana, ideogram. Brief consumers: reels-script (veo-prompt-pack),
tiktok-script, youtube-shorts, youtube-long-form, cross-platform-repurposing.
Publish: scheduling-and-queue → WoopSocial.
Definition of done
The job named; a portable, brand-matched brief (no vendor lock-in); routed to the right tool
category and current tool (verify-quarterly) with prompt craft handed to the mini-skill; "film it
instead" considered; Sora and other dead tools never routed to; AI disclosure planned; publishing
routed to scheduling-and-queue → WoopSocial; no deepfakes, no fabricated metrics.