- Home
- Skills
- Image & Video
- Shorts Production: Human-Gated Faceless Video Workflow
Works with the AI Agents you already use
Shorts Production: Human-Gated Faceless Video Workflow
Topic to finished 1080x1920 Short with any image model, TTS and editor: sourced script, beat plan, consistent characters, storyboard sign-off before any spend, QA and a release package.
$9
Shorts Production: Human-Gated Faceless Video Workflow
Example session with this skill installed
Make me a 45-second history Short about the Great Molasses Flood, gouache style, ElevenLabs voice, I edit in CapCut.
- Read your context and instructions
- Compiled the shorts production
Storyboard ready: "The Day a Wave of Molasses Hit Boston"
16 beats | target 52 s | estimate 52.7 s (+1.3%) | 137 spoken words @ 2.6 w/s | hook 2.7 s | plan hash f13f1157ce79 | validate_plan.py:
VALID
Sources: 7 verified (book-length history of the disaster, Boston newspapers of Jan 1919, a 2016 fluid-dynamics study of the wave). 2 speculation, each tied to a hedged sentence in the narration ("Locals still say...", "The lesson:..."). 1 disputed claim (the company's sabotage defense) left out.
World bible: 1910s gouache editorial style. Recurring cast: a fictional composite foreman (walrus moustache, flat cap, charcoal waistcoat, never a coat) and a dapple-grey dray horse with a white blaze. 5 locations with anchor shots. Continuity rules scoped to beats: the tank is always on the LEFT, intact in beats 3-7, ruptured from beat 8, gone in beat 16. No readable text, no gore, no real victims shown.
| # | t (s) | VO | Shot | Cast | Location |
|---|---|---|---|---|---|
| 1 | 0.0+2.7 | Boston, 1919: a wave of molasses. | low-angle wide | - | commercial_st |
| 2 | 2.7+3.5 | January fifteenth was oddly warm in the North End. | medium-wide | foreman | north_end_row (ANCHOR) |
| 3 | 6.2+3.9 | By the harbor stood a steel tank, five stories tall. | extreme wide | dray_horse | commercial_st (ANCHOR) |
| 4 | 10.0+2.7 | Inside: over two million gallons of molasses. | cutaway, low angle | foreman | tank_base (ANCHOR) |
| 5 | 12.7+2.3 | It had leaked from the start. | close-up | - | tank_base |
| 6 | 15.0+3.9 | The owners painted it brown, so the leaks wouldn't show. | medium | foreman | tank_base |
| ... | (16 rows) |
Prompts: prompts.md with 16 assembled per-beat prompts (each under 220 words, only the continuity rules that apply to that beat).
Open warning: the commercial_st anchor is beat 3 while the location first appears in beat 1 (flash-forward), so beat 3 is generated first.
Generation plan: 16 beats + 2 identity sheets, at most 30 image calls including re-rolls, ~55 s of ElevenLabs narration, CapCut assembly with word-timed captions and music ducking; qa_video.py --timeline after export.
Nothing has been generated or spent yet. Approve this storyboard and a budget ceiling to start, or tell me which beats to change.
Connects securely to your tools. The creator never sees your data.
About this skill
Turn any topic into a finished 1080x1920 Short, Reel or TikTok without burning credits on images you'll throw away, and without an agent publishing something you never saw.
What your agent does with this skill
- Researches the topic and keeps a source ledger: verified facts, hedged speculation (tied to the exact narration sentence), disputed claims left out, and a rule for what counts as independent sources.
- Writes narration sized to your target duration using words-per-second math (pronunciation-aware), with a hook under ~3 s.
- Splits it into 12-20 visual beats whose voice-over is a verbatim slice of the script, cut only at word boundaries, so captions and cuts always line up.
- Builds a world bible: character identity cards, outfits, props, locations with anchor shots, palette, period, banned items, and continuity rules scoped to the beats they apply to.
- Assembles per-beat image prompts that keep characters and places consistent in any model (Gemini, GPT-image, Flux, Midjourney, SDXL) with model-specific notes: reference adapters and token limits for Flux, recommended sizes.
- Stops at Gate 1: a storyboard plus a budget (or a call count for free local GPUs) you approve before any generation.
- Generates with a re-roll decision table and a spend ledger, then TTS, local word timing (whisper.cpp or per-beat fallback), and assembly in ffmpeg, Remotion, CapCut or DaVinci. The ffmpeg path is fully tested: eased zoom/pan/tilt/shake/rotate, stereo voice+music mix with measured ducking, and burned captions even on builds without libass.
- Runs automated QA: decode, resolution, black and frozen frames, LUFS and true peak, mono or short audio, silences on the voice track, caption/narration match, beat lengths. Then a visual QA for cast, duplicate props and stray text.
- Prepares a release package and stops at Gate 2: nothing is uploaded or made public without your explicit go-ahead naming the visibility.
Included
validate_plan.py: placeholders, script/beat drift down to the first differing character, mid-word cuts, duration vs target, hook length, unhedged speculation, broken world references, missing anchors, prohibited items (plural-aware, negation-aware). Writes the storyboard and an approval hash.assemble_prompts.py: builds the per-beat prompt file with only the relevant continuity rules and a word cap.qa_video.py: JSON QA report plus a per-beat contact sheet. Python standard library and FFmpeg only; no network, no API keys.- JSON Schemas, in-depth guides and a complete validated 16-beat example episode with storyboard and publish package.
Buyer-tested end to end on a 40 s Flux + local TTS + ffmpeg run: valid plan, QA-passing mp4.
Where it stops
It doesn't generate images, voice or video itself: it drives the tools you already have. Visual QA (identity, duplicate props, stray text) needs a model that can see images, or you. The agent never spends credits, approves or publishes on its own.
- It does not generate images, voice or video itself; it uses the tools you already have.
- Visual QA (identity, duplicate props, stray text) needs a model that can see images, or a human; the automated checks cover technical defects only.
- Word-level timing with local TTS needs a speech-to-text model (e.g. a whisper.cpp model download); without one the per-beat fallback gives approximate word timing.
- Burned captions in plain ffmpeg need libass; on builds without it the skill's image-overlay fallback needs a way to render caption PNGs (e.g. Pillow).
- Prices, platform limits and model features change, so the skill tells the agent to confirm them rather than hard-coding them. Duration is an estimate until the real voice is measured.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 9 days ago
Requires