hearyourvoice

🎙️ HearYourVOICE

English · ภาษาไทย ↓

The complete, repeatable workflow for short Thai documentary/explainer videos — from a topic to a finished MP4, in whatever format you ship.

HearYourVOICE is an agent skill + a 16-strong subagent team — a SKILL.md, its references/ and scripts/, and the hyv-* specialists that do the work. Drop a topic in, and it walks the whole loop: research → script → an adversarial agent debate over the hook/punchline → ElevenLabs voiceover → gather visuals (self-shot, found Creative-Commons, or generative) → lay timecoded inserts on an editor timeline → validate → package one clean delivery/<slug>/ folder.

The point: the harness, not the model

Stop asking "which LLM is smarter." It's the wrong question.

You are the one in command. You define the format, the process, the standard — and the model follows it. An LLM is just a brain you can swap out: Claude today, something else tomorrow. What stays is the harness — the skill, the scripts, the guardrails that reliably do the work you guide them to do.

A good AI agent isn't the one that wins a chat-window argument about which model is best. It's the one with a harness complete enough to pick up a real task and finish it. If your AI ability ends at chatting and ranking models, you're about two years behind where this is already going.

HearYourVOICE is that harness for video. SKILL.md is production know-how made executable — and any brain in the world can run it.

The team — run it solo, or delegate the whole thing

Run the phases yourself, or hand them to the team that ships with the skill. hyv-producer is the director: it delegates to the specialists, holds the gates, and is the only one that talks to you about them.

hyv-producer  ← the director: delegates, holds gates, never decides taste
  ├─ hyv-researcher       phase 0 · real WebSearch, every claim ≥2 sources
  ├─ hyv-scriptwriter     phase 1 · script + TTS-ready narration
  ├─ hyv-script-reviewer  phase 1 · strict pass/fail against the brief
  ├─ hyv-storyboard       phase 1 · shot-by-shot, at the config's aspect
  ├─ debate panel         phase 1 · hook-maximalist vs skeptic-editor vs target-viewer
  │   └─ hyv-judge          synthesizes · RECOMMENDS, never rules
  ├─ hyv-shotlister       phase 1 · the 5-sheet shotlist.xlsx backbone
  ├─ hyv-voiceover        phase 2 · ElevenLabs + ffprobe = the master clock
  │
  │  ⟨ phase 3 · filming your own shots — no agent here. That part is you. ⟩
  │
  ├─ hyv-cc-scout         phase 4 · finds CC footage, verifies the license itself
  ├─ hyv-veo-prompt-smith phase 4 · prompts for the gaps · spends nothing
  ├─ hyv-veo-runner       phase 4 · renders generative clips · costs money
  └─ hyv-assembler        phase 5–6 · timeline → editor → delivery/

(Phase 3 has no agent on purpose — an AI can't hold your camera. Shoot, cut exact-timestamp selects, drop them in the pool, and hyv-assembler picks them up like any other clip.)

Two rules the team never breaks. Creative authority is human: hyv-judge recommends and flags a split, but you pick the hook and punchline. And every gate stops for you — output spec, topic lock, split debate, and any spend (voiceover or generative credits).

Fan them out to scale: one hyv-researcher per topic in parallel, then one hyv-producer per video — each in its own src/<slug>/, no collisions. Full roster + decision chain: references/subagents.md.

Nothing is hardcoded

Aspect, resolution, fps, and editor live in src/<slug>/project.config.json — confirmed with you before anything is produced. 9:16 is the default, not a law: 16:9, 1:1, 4:5, and custom work the same way, and every script takes --aspect/--width/--height/--fps.

Your editor is a config value too. Phase 5 exports one universal timeline — ep*-timeline.json (frames + SMPTE timecode) and ep*-timeline.csv (one row per clip) — and hyv-assembler reads editor to take the right path:

editor	What happens
capcut · premiere · davinci	You get the CSV/JSON and drop each clip at its start_tc in the NLE you already know. Nothing to install.
remotion	The code path. new-project.mjs --remotion scaffolds a starter composition and config; the timeline JSON drives it directly, and remotion-best-practices loads for the render. Pick this when you want the cut to be reproducible from source, or when you're generating many videos from one template.

The recipe is identical wherever you take it: voiceover on one audio track (the master clock), every clip muted at its start timecode, at the config's aspect and fps. Full recipe: references/assembly-and-validation.md.

The loop — seven phases