skill
video-production-skills
Agent Skills for complete AI video production, from character sheets and storyboards to shot production, editorial, mastering and evaluation.
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/sb-dev/video-production-skills itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 22, 2026Video Production Skills
Direct AI video productions, not isolated generations.
Video Production Skills gives AI coding agents a production workflow for turning a brief into a finished video.
It supports the full production process:
- Creative definition: visual direction, character sheets and manifests, product manifests
- Previsualisation: storyboards, shot plans, animatics
- Shot development: reference frames, motion prototypes, video-shot production and selection
- Editorial: continuity, editing, audio integration
- Finishing and delivery: mastering and delivery variants
- Quality: evaluation, technical QC and targeted refinement.
The workflow is designed to plan cheaply, approve creative decisions, preserve continuity, select shots, edit, evaluate, and retry only what failed.
Approval and cost control
Creative decisions are approved by you, not by the agent.
- Selection is the agent's. It may advance an artifact to
selected— "keep developing this". - Approval is yours. Only you move an artifact to
approvedorlocked, and the approver is recorded inapprovedBy. An artifact marked approved with nobody recorded as approving it is treated as unapproved, andtools/validate-production.tsreports it.
Approval is requested before expensive generation for visual direction, character and identity references, environment references, the shot plan, and picture lock.
The agent asks before inventing a story-changing parameter it was not given — who the characters are, where the piece is set, which props carry meaning — because the cheapest representation that resolves a creative question is a question.
Retries and spend are bounded. When corrective attempts at one layer hit the agreed ceiling, or generation passes the agreed budget, the run stops and reports rather than escalating cost quietly.
Install
npx skills add sb-dev/video-production-skills \
--skill video-production \
--skill video-evaluate \
--agent claude-code
For the default image/video execution path:
npx skills add replicate/skills \
--skill find-models \
--skill compare-models \
--skill prompt-images \
--skill prompt-videos \
--skill run-models \
--agent claude-code
For Codex: use --agent codex instead.
Set REPLICATE_API_TOKEN when generation is required.
Quick start — Mechanical Watch Hero
Start with one shot and learn the core production loop.
Use video-production to create a 6-second luxury product hero shot.
Create a cinematic macro video of a black mechanical wristwatch resting on a brushed metal surface. The shot should feel premium, controlled, and elegant.
Requirements:
- Duration: 6 seconds
- Format: 16:9
- Tone: luxury commercial
- Subject: one black mechanical watch with a detailed dial, polished case, and visible texture on the strap
- Camera: slow, precise push-in with a subtle lateral move
- Lighting: dark studio environment with a narrow moving highlight sweeping across the watch face and case
- Focus: preserve crisp dial details, indices, hands, reflections, and premium material feel
- Background: minimal, dark, unobtrusive
- Motion: no chaotic movement; everything should feel controlled and intentional
Workflow:
- Establish concise visual direction first
- Create a reference frame before final video generation
- Generate at least two shot candidates
- Select the strongest candidate
- Run evaluation and technical QC
What to optimise for:
- premium product fidelity
- reflection control
- readable dial details
- smooth camera motion
- elegant lighting transition
A one-shot project should stay small:
production/video/
├── brief.md
├── references/
├── shots/shot-001/
│ ├── reference-frame.png
│ ├── candidates/
│ ├── selection.json
│ └── provenance.json
├── master/video-master.mp4
└── evaluation/
The important behaviour is:
cheap checkpoint → approve → generate → evaluate → refine only what failed
Learn by producing
Progress through increasingly demanding productions. Each adds a real production problem, not another abstraction.
Level 1 — Control one shot
6s single shots
Learn candidates, selection, refinement, and QC. Each example folder holds the prompt, the generated files, and an explanation of what happened.
| Mechanical Watch Hero — luxury product hero | Midnight Espresso — food/product commercial |
| Last Train Portrait — cinematic character shot | Boxer at Dawn — sports-action shot |
Level 2 — Make shots work together
12–15s three-shot sequences
Learn character continuity, storyboard, shot planning, and editing.
| Last Train Home — three-shot character sequence | Missed Connection — two-character near miss |
| The Red Umbrella — colour-anchor tracking sequence | The Letter — restrained prop drama |
Adds visual-direction.md, storyboard/, shot-plan.json, and edit/.
Level 3 — Produce a commercial
15s commercials with audio and delivery variants
Learn product fidelity, graphics, audio, and delivery variants.
| Obsidian No. 7 — luxury fragrance | Arc One — wireless headphones |
| Apex — running shoe | Ember — espresso machine |
Adds product references/manifests, graphics/, audio/, and delivery/.
Level 4 — Direct a short
30–45s multi-sequence micro-films
Learn shared references, sequences, animatics, and editorial iteration.
| Signal Lost — sci-fi micro-film | Night Courier — urban action micro-film |
| The Last Light — atmospheric storm short | Room 417 — suspense micro-film |
Adds shared/ references and sequences/ when a flat shot list stops scaling.
Level 5 — Run a video campaign
Hero film + product film + social + bumper
Learn multiple productions and cross-domain handoffs.
| Aster — electric motorcycle campaign | Nova — AR glasses campaign |
| Helios — home robot campaign | Atlas One — adventure e-bike campaign |
At this level the project can compose other Creative Production Skills through artifacts:
production/
├── advertising/ # audience, proposition, claims
├── narrative/ # story and scenes
├── music/ # composition and master
└── video/ # visual production and video masters
Project structure grows with the work
One shot
Use brief, references, shots, master, and evaluation.
Shots must work together
Add visual-direction, storyboard, shot-plan, and edit.
Product, audio, or delivery matter
Add manifests, graphics, audio, and delivery.
A flat shot list stops scaling
Add shared and sequences.
Several distinct videos share assets
Use video/shared and video/productions.
Other creative domains contribute
Use production/{advertising,narrative,music,video}.
Keep the structure lean:
- a shot is the smallest retryable production unit;
- selections point to candidates instead of copying them into
selected/folders; - lifecycle state lives in artifact metadata, not global
draft/approved/finaltrees; shared/exists only for genuinely shared artifacts;- masters stay separate from delivery variants.
Skills
video-production
Plan, draft, generate, select, edit and finish video while preserving approved decisions.
video-evaluate
Diagnose the failure layer and recommend the smallest corrective action.
Evaluation is actionable rather than just a score:
wrong composition → revise reference frame or affected shot
wrong camera movement → revise shot plan and affected shot
good shots, poor pace → revise edit
corrupt media → fail technical QC
video-extension-pack-creator
Author reusable extension packs — a format, genre, style, audience and optional voice casting held together as one production profile — and validate and catalogue them before they ship.
A pack changes the quality target, not the requirement for quality. It ranks below explicit user instruction and approved artifacts.
Execution
The skills decide what production work is needed. Existing tools execute it.
- Replicate Agent Skills — default image/video model discovery and execution
- ElevenLabs Agent Skills — optional speech, transcription and sound effects
- FFmpeg / ffprobe — assembly, inspection and QC
- ImageMagick — deterministic image operations
- fal.ai / genmedia — optional secondary execution path
Deterministic scripts are TypeScript and target Node.js 24.12+.
Use the least expensive representation that can resolve the current production uncertainty.
Preserve approved decisions and change only what needs to change.
Documentation
Specifications
- Creative Skills System Specification
- Creative Skills Workflows and Artifacts Specification
- Creative Skills Repository and Contracts Specification
- Testing and Benchmark Specification
- Video Customisation Packs Specification
- Video Extension Pack Catalogue Specification
Process and reports
Skills
Evals
Benchmarks
- Benchmark runbook — suites,
npm run benchmark:list, preparing and scoring cases, coverage gates
Examples
- Level 1 — Control one shot
- Level 2 — Make shots work together
- Level 3 — Produce a commercial
- Level 4 — Direct a short
- Level 5 — Run a video campaign
Contributing
See CONTRIBUTING.md.
License
Apache-2.0. See LICENSE.