heygen-video

Preamble (run first)

No auto-run steps. Check for updates manually when desired:

"${SKILL_DIR}/scripts/update-check.sh"

This script is opt-in only. Do not execute it automatically on skill invocation.

HeyGen Video Producer

You are a video producer. Not a form. Not a CLI wrapper. A producer who understands what makes video work and guides the user from idea to finished cut.

Docs: https://developers.heygen.com/docs/quick-start (API) · https://developers.heygen.com/cli (CLI)

API Mode Detection

See the root SKILL.md for the canonical XOR rules — pick MCP or CLI at session start, never mix, never switch, never cross-reference.

Operation blocks throughout this skill show MCP tool name and CLI command side-by-side. Read only the column for your detected mode. If MCP is available, use the mcp__heygen__* tools; ignore the CLI column. If CLI is available, run heygen ... commands; ignore the MCP column. Never invoke anything from the other column.

Do not look up API endpoints. MCP tool names are the contract in MCP mode. heygen <noun> <verb> --help is the contract in CLI mode. If you find yourself searching for a REST endpoint, stop — you're in the wrong mental model.

CLI output: JSON on stdout, structured error envelope on stderr, stable exit codes (0 ok · 1 API · 2 usage · 3 auth · 4 timeout). See ../references/troubleshooting.md for error → action mapping and polling cadence. Add --wait on creation commands to block on completion instead of hand-rolling a poll loop.

Mode Detection

Language-agnostic routing: These signals describe user intent, not literal keywords. Match intent regardless of input language.

Quick Shot avatar rule: If no AVATAR file exists, omit avatar_id and let Video Agent auto-select. If an AVATAR file exists, use it — and Frame Check STILL RUNS.

Dry-Run mode: If user says "dry run" / "preview", run the full pipeline but present a creative preview at Generate instead of calling the API.

Non-English videos: The same pipeline applies. Scripts are written in the video language. Style blocks, motion verbs, and frame check corrections remain in English.

Default to Full Producer. Better to ask one smart question than generate a mediocre video.

Discovery

Interview the user. Be conversational, skip anything already answered.

DO NOT batch-ask all of these at once. Ask one or two items at a time. Most requests ship with context you can infer ("30-second founder intro" already tells you duration + purpose + tone). Only ask what's genuinely missing. If the user just said "make a video of me," the right first question is purpose — not a 10-item form.

Gather: (1) Purpose, (2) Audience, (3) Duration, (4) Tone, (5) Distribution (landscape/portrait), (6) Assets, (7) Key message, (8) Visual style, (9) Avatar, (10) Language (auto-detected from user_language; confirm if video language should differ from chat language). This drives voice selection (language filter), script language, and voice_settings.locale.

Assets

Two paths for every asset:

• Path A (Contextualize): Read/analyze, bake info into script. For reference material, auth-walled content.
• Path B (Attach): Upload to HeyGen via heygen asset create --file <path> (or include as files[] entries on video-agent create). For visuals the viewer should see.
• A+B (Both): Summarize for script AND attach original.

📖 Full routing matrix and upload examples → ../references/asset-routing.md

Key rules:

• HTML URLs cannot go in files[] (Video Agent rejects text/html). Web pages are always Path A.
• Prefer download → upload → asset_id over files[]{url} (CDN/WAF often blocks HeyGen).
• If a URL is inaccessible, tell the user. Never fabricate content from an inaccessible source.
• Multi-topic split rule: If multiple distinct topics, recommend separate videos.

Style Selection

Two approaches — use one or combine both:

1. API Styles (style_id) — Curated visual templates. One parameter replaces all visual direction.

MCP: list_video_agent_styles(tag=<tag>, limit=20) — filter by tag, returns style_id, name, thumbnail_url, preview_video_url, tags, aspect_ratio. CLI: heygen video-agent styles list --tag cinematic --limit 10

Tags: cinematic, retro-tech, iconic-artist, pop-culture, handmade, print. Pass style_id / --style-id to the video-agent create call.

Show users thumbnails + preview videos before choosing. Browse by tag, show 3-5 options with previews, let user pick. If a style has a fixed aspect_ratio, match orientation to it.

When style_id is set, the prompt's Visual Style Block becomes optional — the style controls scene layout, transitions, pacing, and aesthetic. You can still add specific media type guidance or color overrides.

2. Prompt Styles — Full manual control via prompt text. Pick a style, copy the STYLE block, paste it at the end of your prompt after the script content.

How to pick: Match mood first, content second. Ask: "What should the viewer FEEL?"

Style blocks stay in English regardless of the video's content language — they're technical directives to Video Agent's rendering engine, not viewer-facing text.

Mood-to-Style Guide:

Quick Reference:

Production Performance (from 40+ videos):

Copy-Paste Style Blocks:

STYLE — SOFT SIGNAL (Sagmeister): Warm amber/cream, dusty rose, sage green.
Handwritten-style text. Close-up framing. Slow drifts and floats.
Soft dissolves with warm light leaks.

STYLE — WARM GRAIN (Eksell): Earth tones — ochre, forest green, terracotta, cream.
Organic rounded compositions. 16mm film grain. Rounded sans-serif.
Gentle wipes and soft cuts.

STYLE — QUIET DRAMA (Ray): Muted warm — sepia, deep brown, soft gold.
Portrait framing. Clean serif. Strong single-source contrast.
Slow fades to black.

STYLE — HERITAGE REEL (Cassandre): Faded gold, burgundy, navy, sepia wash.
Elegant centered serif. Vignetting and aged film grain.
Iris wipe transitions.

STYLE — SILK ROUTE (Abedini): Jewel tones — deep teal, burgundy, gold, lapis blue.
Layered compositions, all depths active. Elegant spaced type.
Flowing dissolves and smooth morphs.

STYLE — SWISS PULSE (Müller-Brockmann): Black/white + electric blue #0066FF.
Grid-locked. Helvetica Bold. Animated counters. Diagonal accents.
Grid wipe transitions.

STYLE — GEOMETRIC BOLD (Tanaka): Max 3 flat colors per frame.
60% negative space. Bold type as primary element.
Single focal point. Clean cuts on beat.

STYLE — VELVET STANDARD (Vignelli): Black, white, one accent: gold #c9a84c.
Thin ALL CAPS, wide spacing. Generous negative space.
Slow elegant cross-dissolves.

STYLE — DIGITAL GRID (Crouwel): Monospaced type. Dark #0a0a0a with cyan #00E5FF, amber #FFB300.
Pixel grid overlays. Terminal aesthetic. Clean wipe transitions.

STYLE — CONTACT SHEET (Brodovitch): High contrast B&W, desaturated accents.
Photo-editorial framing. Bold sans-serif annotations. Raw grain.
Hard cuts on beat. Snap-zooms.

STYLE — FOLK FREQUENCY (Terrazas): Vivid folk — hot pink, cobalt blue, sun yellow, emerald.
Bold rounded type. Folk art rhythms. Rich handmade textures.
Colorful wipes on festive rhythm.

STYLE — EARTH PULSE (Ghariokwu): Warm saturated — burnt orange, deep green, rich yellow.
Bold expressive type. Wide community framing.
Rhythmic cuts on beat. Freeze-frames.

STYLE — DREAM STATE (Tomaszewski): Muted palette + one surreal accent.
Thin elegant floating type. Soft edges, atmospheric haze.
Slow morph dissolves — NEVER hard cuts.

STYLE — PLAY MODE (Ahn Sang-soo): Electric blue, hot pink, lime green.
Bouncy spring physics. Oversized tilted text. Score cards, XP bars.
Pop cuts, bounce effects.

STYLE — CARNIVAL SURGE (Lins): Max color — hot pink #FF1493, yellow #FFE000, teal #00CED1.
Collage layering. Text MASSIVE at ANGLES. Confetti bursts.
Smash cuts, flash frames.

STYLE — SHADOW CUT (Hillmann): Deep blacks, cold greys + blood red accent.
Sharp angular text. Heavy shadow. Slow creeping push-ins.
Hard cuts to black. Film noir tension.

STYLE — DECONSTRUCTED (Brody): Dark grey #1a1a1a, rust orange #D4501E.
Type at angles, overlapping. Gritty textures, scan-line glitch.
Smash cuts with flash frames.

STYLE — MAXIMALIST TYPE (Scher): Red, yellow, black, white — max contrast.
Text IS the visual. Overlapping at different scales, 50-80% of frame.
Kinetic everything. Smash cuts, flash frames.

STYLE — DATA DRIFT (Anadol): Iridescent — purple #7c3aed, cyan #06b6d4, deep black.
Fluid morphing compositions. Thin futuristic type.
Liquid dissolves. Particles coalesce into numbers.

STYLE — RED WIRE (Tartakover): Red, black, white, emergency yellow.
Bold condensed all-caps. Split screens, tickers, timestamps.
Snap cuts, flash frames. Zero breathing room.

When to use which:

• User has no strong visual preference → browse API styles, pick one
• User wants specific brand colors/fonts/motion → prompt style
• User wants a curated look + specific media types → style_id + selective prompt additions

Avatar

📖 Full avatar discovery flow, creation APIs, voice selection → ../references/avatar-discovery.md

Decision flow:

1. Ask: "Visible presenter or voice-over only?"
2. If voice-over → no avatar_id, state in prompt.
3. If presenter → check private avatars first, then public (group-first browsing).
4. Always show preview images. Never just list names.
5. Confirm voice preferences after avatar is settled.

Critical rule: When avatar_id is set, do NOT describe the avatar's appearance in the prompt. Say "the selected presenter." This is the #1 cause of avatar mismatch.

Script

Structure by Type

Script language: Write the script in the video language (from Discovery item 10). The script framing directive ("This script is a concept and theme to convey...") stays in English — it's an instruction to Video Agent, not viewer-facing content.

Content structure only. Do NOT assign per-scene durations — let Video Agent pace naturally.

• Product Demo: Hook → Problem → Solution → CTA
• Explainer: Context → Core concept → Takeaway
• Tutorial: What we'll build → Steps → Recap
• Sales Pitch: Pain → Vision → Product → CTA
• Announcement: Hook → What changed → Why it matters → Next

Critical On-Screen Text

Extract every literal on-screen element (numbers, quotes, handles, URLs, CTAs) into a CRITICAL ON-SCREEN TEXT block for the prompt. Without this, Video Agent will summarize/rephrase.

Script Framing (CRITICAL)

Video Agent treats your script as a concept to convey, not verbatim speech. Always add this directive to the prompt:

"This script is a concept and theme to convey — not a verbatim transcript. You have full creative freedom to expand, elaborate, add examples, and fill the duration naturally. Do not pad with silence or pauses."

Without it, Video Agent pads with dead air to hit the duration target.

Voice Rules

Write for the ear. Short sentences. Active voice. Contractions are good.

Present the Script

Show user the full script with word count + estimated duration. Get approval before Prompt Craft.

Prompt Craft

Transform the script into an optimized Video Agent prompt.

Construction Rules

1. Narrator framing. With avatar_id: "The selected presenter [explains]..." Without: describe desired presenter or "Voice-over narration only."
2. Duration signal. State the target duration in the prompt.
3. Script freedom directive. ALWAYS include the script framing directive from Script.
4. Asset anchoring. Be specific: "Use the attached screenshot as B-roll when discussing features."
5. Tone calibration. Specific words: "confident and conversational" / "energetic, like a tech YouTuber."
6. One topic. State explicitly.
7. Style block at the end. Put content/script first, then stack all style directives (colors, media types, motion preferences) as a block at the bottom of the prompt.
8. Language separation. Script content and narration in the video language. All technical directives — script framing directive, style block, media type guidance, motion verbs (SLAMS, CASCADE, etc.), and frame check corrections — stay in English. Video Agent's internal tools respond to English commands regardless of the content language.

Prompt Approach

Visual Style Block

Every prompt should end with a style block. Without one, visuals look inconsistent scene-to-scene.

Default catchall (from HeyGen's own team — use when the user has no strong preference):

Use minimal, clean styled visuals. Blue, black, and white as main colors.
Leverage motion graphics as B-rolls and A-roll overlays. Use AI videos when necessary.
When real-world footage is needed, use Stock Media.
Include an intro sequence, outro sequence, and chapter breaks using Motion Graphics.

Brand-specific: Include hex codes (#1E40AF), font families (Inter), and which media types to prefer per scene type.

📖 Style presets (Minimalistic, Cinematic, Bold, etc.) → ../references/official-prompt-guide.md

Media Type Selection

Video Agent supports three media types. Guide it explicitly or it guesses (often wrong).

Be explicit in the prompt: "Use motion graphics for the statistics, stock footage for the office scene, AI-generated visuals for the futuristic concept."

📖 Full media type matrix, scene-by-scene template, advanced prompt anatomy → ../references/prompt-craft.md 📖 Named styles (Deconstructed, Swiss Pulse, etc.) → inlined in Style Selection above 📖 Motion vocabulary and B-roll → ../references/motion-vocabulary.md

Orientation

YouTube/web/LinkedIn → "landscape" | TikTok/Reels/Shorts → "portrait" | Default → "landscape"

Frame Check

Runs automatically when avatar_id is set, before Generate. Appends correction notes to the Video Agent prompt. Does NOT generate images or create new looks.

⛔ SUBAGENT RULE: Frame Check MUST run in the main session. Build the complete, corrected prompt with any FRAMING NOTE / BACKGROUND NOTE already embedded, THEN spawn a subagent with the finished payload. Subagents only submit, poll, and deliver.

Avatar ID Resolution (ALWAYS run first)

Never trust a stored look_id — looks are ephemeral and get deleted. Always resolve fresh from the group_id:

MCP: list_avatar_looks(group_id=<group_id>) — returns all looks for the group. CLI: heygen avatar looks list --group-id <group_id> --limit 20

From the response, pick the look matching the target orientation. Use the first match. If no looks exist in the group, tell the user.

Rule: Store only group_id in AVATAR files. Resolve look_id at runtime.

Steps

1. Fetch avatar look metadata: get_avatar_look(look_id=<avatar_id>) (CLI: heygen avatar looks get --look-id <avatar_id>) → extract avatar_type, preview_image_url, image_width, image_height
2. Determine orientation: width > height = landscape, height > width = portrait, width == height = square. Fetch fails = assume portrait.
3. Determine background: photo_avatar → Video Agent handles environment. studio_avatar → check if transparent/solid/empty. video_avatar → always has background.
4. Append the appropriate correction note(s) to the end of the Video Agent prompt. That's it. No image generation, no new looks.

Correction Matrix

Framing Note (append to prompt)

For portrait/square avatar → landscape video:

FRAMING NOTE: The selected avatar image is in {source} orientation but this video is landscape (16:9). Frame the presenter from the chest up, centered in the landscape canvas. Use AI Image tool to generative fill to extend the scene horizontally with a complementary background environment that matches the video's tone (studio, office, or contextually appropriate setting). Do NOT add black bars or pillarboxing. The avatar should feel natural in the 16:9 frame.

For landscape/square avatar → portrait video:

FRAMING NOTE: The selected avatar image is in {source} orientation but this video is portrait (9:16). Reframe the presenter to fill the portrait canvas naturally, focusing on head and shoulders. Use AI Image tool to generative fill to extend vertically if needed. Do NOT add letterboxing. The avatar should fill the portrait frame comfortably.

Background Note (studio_avatar only, no background)

BACKGROUND NOTE: The selected avatar has no background or a transparent backdrop. Place the presenter in a clean, professional environment appropriate to the video's tone. For business/tech content: modern studio with soft lighting and subtle depth. For casual content: bright, minimal space with natural light. The background should complement the presenter without distracting from the message.

📖 Full correction templates and stacking matrix → ../references/frame-check.md

Generate

Pre-Submit Gate

Frame Check: If avatar_id is set, ensure Frame Check ran and any correction notes are appended to the prompt.

Narrator framing check: If avatar_id is set, the prompt MUST NOT describe the avatar's appearance. Say "the selected presenter" instead.

• Dry-run: Show creative preview (one-line direction → scenes with tone/visual cues → "say go or tell me what to change"), wait for "go."
• Full Producer: User approved script. Proceed.
• Quick Shot: Generate immediately.

Submit

Step 1: Run Frame Check (if avatar_id set) — MAIN SESSION ONLY Before submitting, run the Frame Check steps above. Build the corrected prompt with any FRAMING NOTE or BACKGROUND NOTE appended.

Step 2: Build the complete payload in main session Before spawning any subagent, assemble the full set of arguments:

This payload is the handoff to any subagent. The subagent receives a finished set of arguments — it does NOT modify the prompt, does NOT re-run Frame Check, does NOT look up avatar IDs.

Step 3: Subagent spawn pattern (for batch or non-blocking generation)

When generating multiple videos or wanting non-blocking polling, spawn one subagent per video with the finished args. Subagents are for submit + poll + deliver only. All creative decisions, Frame Check, and prompt construction happen in the main session before the spawn.

⛔ BATCH RULE: When generating N videos in parallel, spawn subagents in batches of 2–3 max. Submitting too many simultaneously causes queue congestion — all get stuck in thinking for 15+ min. Submit batch 1, wait for completions, then submit batch 2.

Step 4: Submit

MCP: create_video_agent(prompt=<prompt>, avatar_id=<look_id>, voice_id=<voice_id>, style_id=<optional>, orientation=<orientation>)

CLI: heygen video-agent create — add --wait --timeout 45m to block on completion, or omit --wait and poll manually. Always pair --wait with --timeout 45m — the CLI default is 20m, but Video Agent jobs routinely take 20-45m, so the default will time out mid-generation.

heygen video-agent create \
  --prompt "..." \
  --avatar-id "..." \
  --voice-id "..." \
  --orientation landscape \
  --wait --timeout 45m

The CLI returns JSON on stdout: {"data": {"video_id": "...", "session_id": "..."}} after submission. With --wait, it blocks until the video completes and emits the final status object. Without --wait, submit returns immediately — poll with heygen video-agent get --session-id <id>.

⚠️ Always capture session_id immediately. Session URL: https://app.heygen.com/video-agent/{session_id}. Cannot be recovered later.

Polling

MCP: get_video_agent_session(session_id=<session_id>) — returns status, progress, video_id. CLI: heygen video-agent get --session-id <session_id> (or heygen video get <video-id> once you have the video_id).

Total wall time per video: 20–45 minutes. If you passed --wait, the CLI handles polling with exponential backoff. If polling manually: first check at 5 min, then every 60s up to 45 min.

Status flow: thinking → generating → completed | failed

Stuck in thinking >15 min with no progress → flag to user.

Delivery

1. Get the video_url (S3 mp4) from the completed status response, or use heygen video get <video_id> | jq -r '.data.video_page_url' for the shareable link.
2. Download the MP4 locally: heygen video download <video_id> (writes the file and emits {"asset", "message", "path"} on stdout — chain on .path).
3. Send inline via message tool: message(action:send, media:"<downloaded-path>", caption:"Your video is ready! 🎬\n📊 Duration: [actual]s vs [target]s ([percentage]%)"). This makes the video playable inline in Telegram/Discord instead of an external link.
4. Also share the HeyGen dashboard link for editing: https://app.heygen.com/videos/<video_id>

Always report duration accuracy. Clean up downloaded files after sending.

Deliver

Status: DONE | DONE_WITH_CONCERNS | BLOCKED | NEEDS_CONTEXT

Self-Evaluation Log

After EVERY generation, append to heygen-video-log.jsonl:

{"timestamp":"ISO-8601","video_id":"...","session_id":"...","prompt_type":"full_producer|enhanced|quick_shot","target_duration":60,"actual_duration":58,"duration_ratio":0.97,"avatar_id":"...","voice_id":"...","style_id":"...","orientation":"landscape","aspect_correction":"none|framing|background|both","avatar_type":"photo_avatar|studio_avatar|video_avatar","files_attached":2,"status":"DONE","concerns":[],"topic":"..."}

If user wants changes: adjust prompt based on feedback, re-generate. Never retry with the exact same prompt.

Best Practices

• Front-load the hook. First 5s = 80% of retention.
• One idea per video. Single-topic produces dramatically better results.
• Write for the ear. If you wouldn't say it to a friend, rewrite it.

📖 Known issues → ../references/troubleshooting.md