| name | claude-hedra |
| description | Hedra Character-3 lip-synced talking-head video for the AgentCompanies
portfolio. Trigger when a video needs a real character delivering VO
on-camera — intro hooks, presenter beats, outro CTAs, anywhere a viewer
benefits from seeing a face speak the line. Generates from a single
portrait reference + an audio clip; ~$0.50-1.00 per 8s clip.
When NOT to use this skill:
- Atmospheric / cinematic scenes without a named character → fal Sora 2 / Kling 3 Pro
- Animated info-design (charts, timelines, money-flow) → claude-remotion
- Real political figures (Trump, Obama, Adelson, etc.) → Wikimedia photo + ken-burns
Canonical CR narrator portrait: companies/campaign-receipts/brand/betsy-portrait.png
|
claude-hedra — Lip-synced talking-head pipeline
Hedra Character-3 (API: https://api.hedra.com/web-app/public) generates
~8-second lip-synced talking-head clips from a single portrait reference
- an audio file. This is the "viral production value" upgrade — it's what
makes CR look like a show with a host, not a slideshow with AI voice.
When to trigger
A storyboard clip should use Hedra (vendor: "hedra-character3") when:
- A named recurring character delivers a VO line on-camera
- The viewer's retention depends on seeing a face (first 8 seconds of
any video — Hedra intro hook is non-negotiable for the CR channel)
- The line is a "pivot moment" — donor reveal, verdict announcement,
CTA — that needs more weight than a static photo + VO can provide
For CR specifically, the four canonical Hedra beats per long-form:
- Intro hook (0:00-0:08) — Betsy delivers the headline shock
- Mid-1 (~30-40% mark) — Betsy reveals the donor / culprit
- Mid-2 (~60-70% mark) — Betsy delivers the verdict line
- Outro CTA (last 5-8s) — Betsy points viewer to sealed2016.com
API contract (driver invocation)
The produce-from-storyboard.py driver supports vendor: "hedra-character3"
storyboard clips with these required fields:
{
"id": "s1-01-hook",
"vendor": "hedra-character3",
"duration_s": 8,
"portrait": "companies/campaign-receipts/brand/betsy-portrait.png",
"audio": "_build/<slug>/vo-hook.mp3",
"aspect": "16:9"
}
The audio chunk must be pre-extracted from the long-form vo.mp3 via
ffmpeg so Hedra has just the VO segment for this beat (Hedra derives
output duration from audio length).
Standalone CLI
python3 companies/campaign-receipts/scripts/pipeline/hedra-character3.py \
--portrait companies/campaign-receipts/brand/betsy-portrait.png \
--audio _build/<slug>/vo-hook.mp3 \
--duration 8 \
--out _build/<slug>/clips/betsy-hook.mp4 \
--aspect 16:9
Writes the mp4 + a .hedra.json sidecar with the generation_id for cost
tracking.
Auth
Reads HEDRA_API_KEY from /Applications/DrAntoniou Projects/AgentCompanies/.env.
Falls back to legacy NT_HEDRA_API_KEY (same key, shared portfolio-wide).
Never commit the key — .env is gitignored.
House-style rules (CR brand)
- Portrait is canonical. For CR, always use
brand/betsy-portrait.png.
Never re-roll a new portrait per video. Brand consistency across all
145 promise videos depends on Betsy looking like the same person every
time.
- No campaign-rally aesthetic. Hedra's text_prompt should describe
subtle delivery: "warm conversational, slight head movement, eye
contact, natural micro-expressions." Avoid: theatrical gestures,
dramatic emphasis, anchor-stiff posture.
- Speed matches storytelling cadence. The audio chunk should be
ElevenLabs Bella at teacher-cadence settings (stability=0.55,
similarity=0.75, style=0.15, speed=0.93).
Cost discipline
- ~$0.50-1.00 per 8s clip (Hedra credits, rough estimate)
- Per long-form video: 4 Hedra beats × $0.80 = ~$3.20
- Per promise video budget: ≤$5 Hedra spend (the other clips are free
Remotion + ken-burns Wikimedia)
- Hard cap: $13/promise video total
- $500/company cumulative cap (portfolio rule)
Skill location
Canonical: shared/skills/claude-hedra/ in the monorepo (travels with
the repo).
Mirror: ~/.claude/skills/claude-hedra/ for harness auto-discovery.
If you edit the skill, edit the repo copy and cp -R to ~/.claude/skills/.
