agentsope/SkillAlchemy / skills/agentsope-ai-video-director
agentsope-ai-video-director
UnverifiedDirect and implement programmatic videos from a natural-language sentence: brief, shots, style, timeline, assets, engine-specific coding prompts, draft, review and repair. Use when asked to make a video, product launch, explainer, Reel/Short, kinetic typography or docs-to-video with HyperFrames, Remotion or Motion Canvas, or to improve an existing code-video project. 一句话视频制作与导演工作流。
SKILL.md
View on GitHub---
name: agentsope-ai-video-director
description: >-
Direct and implement programmatic videos from a natural-language sentence: brief,
shots, style, timeline, assets, engine-specific coding prompts, draft, review and
repair. Use when asked to make a video, product launch, explainer, Reel/Short,
kinetic typography or docs-to-video with HyperFrames, Remotion or Motion Canvas,
or to improve an existing code-video project. 一句话视频制作与导演工作流。
---
# AI Video Director
## Activation Rules
Use this skill for an authorized video-production task, including:
- “用一句话做一个 15 秒产品发布视频。”
- “Turn this data into a 30-second animated explainer.”
- “Make a narrated video from these authorized docs.”
- “Write the coding-agent prompt and render this Remotion/HyperFrames video.”
- “The captions overlap in this code-video project; inspect and repair it.”
Do not activate solely for:
- A general question about cinema history or a filmmaker.
- A request to summarize a document with no video deliverable.
- A still-image request with no video intent.
- An unrelated React/Vite bug without a video composition.
- A diffusion-model prompt request alone: that extension has not been distilled.
## Agentic Protocol
Execute the following authored application chain (C01–C09).
Every step produces a retained result before its dependent step starts.
Read [operation cards](references/sop_models.md) and match each operation’s exact scope before use.
Do not turn a source-local case into a rule for a new film.
### 1. Turn the sentence into a brief
Read [prompt fields](references/prompt-patterns.md).
Extract goal, audience, main message, platform, supplied facts, duration/aspect,
language, audio intent, assets, engine preference and delivery requirements.
Preserve explicit values. Mark inferred audience, style and format as assumptions.
If unstated, use these reversible **package defaults**, subject to user context:
15s for a short product/social draft; 30s for a compact data explanation;
60s for a compact docs/knowledge draft; 30fps; 16:9 for general explainers,
9:16 when vertical social is explicit; original geometric visuals and system fonts.
These values are authored conventions, not evidence-proven creative optima.
Retain supplied facts with their source. Record unknown facts and synthetic placeholders separately.
A missing product metric must not become a fabricated benefit claim.
Clarify only critical facts, permissions or assets that prevent the requested result;
continue a provisional brief and authorized work while those inputs are pending.
Output `director_brief` with `declared`, `assumptions`, `facts`, `unknowns` and `audio_intent`.
### 2. Make the engine-neutral director plan
Read the matching [video guide](references/video-types.md) and
[director fields](references/director-playbook.md).
Choose an authored communicative pattern from the goal and available material.
Write a shot table: ID, interval in seconds, subject/focal action, framing/layout,
exact copy, animation intent, transition, asset, audio/caption cue and acceptance.
Close the timeline at the declared duration; record intentional overlaps or gaps.
Define palette, typography, hierarchy, backgrounds, safe-area assumptions and motion intent.
Build an asset ledger with provenance, permission, intended use and unresolved rights.
Keep original synthetic visuals available when third-party assets are pending.
For voice-led work, use real authorized audio timing when ready (P28);
if audio is pending, label the entire scene/caption clock provisional.
Do not apply P05’s 2–10-minute or P06’s 90–180-second guidance to short examples as proven rules.
P02/P05/P06/P09 are optional author suggestions, 待验证; they are never quality guarantees.
Output `director_plan`: brief, shots, style tokens, timeline, assets, audio, target and acceptance.
### 3. Probe and route the environment
Read [isolated adapters](references/engine-adapters.md).
Probe actual Node, browser, FFmpeg/exporter, installed versions and local scripts;
record command output/version/license evidence and missing capabilities.
Use an existing compatible project when available; do not silently upgrade it.
Select HyperFrames only within the verified 0.8.142 profile and its local dependencies.
Select Remotion server only with matched 4.0.534 packages, browser and intended-use eligibility (P37).
Its browser compositor is a separate route with its own supported DOM/media/codec subset (P23).
Motion Canvas 3.17.2 uses its Vite editor and registered FFmpeg exporter (P25),
not a fabricated headless render CLI. Preserve release/main and binary-license distinctions (P39).
The Canvas/WebGL renderer is only the recorded source-local profile (P27),
not a universal fallback to recreate or vendor.
Do not buy services or assets. If production rights are unknown, retain the plan/prompt,
use only eligible authorized work and report the unresolved eligibility.
If no runnable eligible exporter exists, output its concrete gap and the achieved state.
Output `engine_route`: engine/profile/version, environment evidence, eligibility and gaps.
### 4. Compile one engine-specific coding prompt
Read the selected adapter’s exact contract and relevant
[HyperFrames example](examples/hyperframes-prompt.md) or
[Remotion example](examples/remotion-prompt.md).
Include the complete director brief, shot table, style, timeline, asset ledger,
audio/caption timing, dependency pins, file/output requirements and acceptance.
Include only the selected adapter’s admitted API/CLI bindings.
For HyperFrames, seconds, static finite root duration and completed paused registration apply (P16–P19).
For Remotion, frame-derived state, zero-based local Sequence clocks and explicit clamps apply (P21).
Audio fields use frames; Caption fields use milliseconds (P40).
For Motion Canvas, generator `yield*` timing uses seconds and persisted event metadata (P25/P31).
Never transpose APIs or clock units between engines.
Specify an actual draft, relevant static/render checks, repair and state report as outputs.
Output `engine_prompt` with the engine identity at its start; never join adapter templates.
### 5. Implement and produce the actual draft
When the chosen environment permits, create original project files and assets,
execute its admitted preview/render path and retain logs, exit codes and output paths.
Do not stop at a prompt when implementation/rendering is authorized and feasible.
If a dependency, critical fact or right blocks a step, preserve the implemented work,
state the exact blocker, and deliver the reachable plan/prompt/code result.
Check the resulting artifact exists; a successful-looking log is not the video.
Output code and an encoded draft only when those artifacts were actually created.
Keep `plan_available`, `code_available` and `encoded_artifact_available` independent.
### 6. Inspect the achieved result
Read [quality rubric](references/quality-rubric.md) and
[checklist](tests/quality-checklist.md).
Check narrative, visual consistency, pacing, animation, legibility, facts, audio,
technical correctness and reproducibility against the declared acceptance.
Run static syntax/resource/timeline checks that the project actually supports.
For encoded drafts, inspect technical metadata, representative and consecutive frames,
then watch full motion and listen when the tools permit; record each coverage separately.
A sampled frame cannot establish continuous pacing or audible synchronization (P35).
For narrated intent without audio, mark audio failed/pending; do not mark listening unnecessary.
Record pass/fail/warning/unknown/not_executed/not_applicable with evidence per dimension.
Output `qa_report`; call a result reviewed only for checks actually performed.
### 7. Repair and hand off truthfully
Turn observed defects into specific shot/code/asset/timing changes.
Use candidate recovery only when its card contains an admitted recovery within scope;
otherwise apply the explicit authored acceptance requirement and label its origin C06.
Rerun affected static/render/frame/motion/audio checks after the change.
Preserve the initial defect, change and new evidence; do not silently replace a failed verdict.
Deliver the current video/code/plan paths, resolved assumptions, sources/rights,
engine/version, commands used, review coverage and remaining gaps.
Record all independent artifact states in the handoff, including preview, encode,
technical probe, sampled frames, full-motion review, listening and reproducibility.
An unexecuted check stays unexecuted even when a different check passed.
Use [manual review](tests/manual-review.md) and its output record contract for evaluation.
## Core Operation Models
General means reusable **inside its exact scope**, not universal or visually validated.
Full cards preserve 40 admissions and every absent recovery/verification component.
| Model | Use within its card’s scope | Primary action |
|---|---|---|
| H1 / P02 | Optional HyperFrames prompt control | Choose delegated aesthetics or explicit layout/timing |
| H2 / P05–P06 | Optional long/structured explainers | Word-led question or scene-frame budget |
| H3 / P09 | Optional Lemo dense sonification/occlusion | Style-local note/pen adjustments |
| H4 / P15 | TikTok auction ad placement only | Match its actual overlay template |
| M1 / P16–P19 | HyperFrames 0.8.142 GSAP profile | Seconds, root/IDs, readiness, ownership |
| M2 / P21–P24 | Remotion 4.0.534 isolated routes | Frame state, export, resource handles |
| M3 / P25–P26/P31 | Released Motion Canvas 3.17.2 | Editor exporter, audio, event metadata |
| M4 / P28/P40 | Ready/pending voice; typed units | Provisional audio or measured clock translation |
| M5 / P37–P38 | Intended use and source-derived assets | Resolve engine and individual asset permissions |
Scoped P01/P03–P04/P07–P08/P10–P14/P20/P27/P29–P30/P32–P36/P39
are conditional source notes in the cards and [case ledger](references/case-studies.md).
Read their source/version/condition before consulting a note; no scoped aesthetic is a generic prescription.
Authored C01–C09 define this package’s workflow/schema/evaluation; empirical evidence IDs are empty for them.
Rule→finding→source/version trace and exact bindings are in [provenance](references/provenance.json).
## Output Style
Lead with the concrete result and actual artifact state.
Use Chinese or the user’s language for useful directions; retain exact API names and units.
Use concise prose plus shot/timeline tables when they help review the film.
Mark assumptions, supplied facts and unknowns explicitly; never present a placeholder as fact.
A coding prompt includes implementable detail, not merely mood words.
Forbidden boilerplate: “Analyzing this with the framework…”, “Following the model card…”,
“Let me analyze this systematically…”, “Would you like me to expand further?”
Cite the original source naturally; keep operation IDs in trace metadata, not promotional claims.
Routine reversible implementation choices do not require an extra approval milestone.
## Output Modes
| Mode | Concrete deliverable | State rule |
|---|---|---|
| Director plan | Brief, shots, style, timeline, assets | Plan available; no render implied |
| Engine prompt | Isolated versioned coding-agent prompt | Prompt available; no code implied |
| Implementation | Original code/assets and execution log | State actual files and blocked steps |
| Rendered draft / QA | Encoded artifact and covered checks | Encoding and review remain independent |
| Repair | Defect, code/timing change, rerun evidence | Only affected checked states advance |
| Handoff | Paths, states, commands, rights, gaps | Unknown/unexecuted states remain visible |
## Boundary Rules
1. Scope: coding-agent programmatic video; diffusion remains a separate undistilled extension.
2. Sources: only admitted P01–P40 within exact scopes plus explicit authored C01–C09;
research cases and structural exemplars never supply additional procedures.
3. Evidence: P02/P05/P06/P09 are optional unverified author guidance;
source claims, code inspection, sampled observation and audiovisual validation differ.
4. Versions: technical knowledge snapshot 2026-10-09; verify exact pins/revision/license
before applying another version. No blanket equivalence from a version string.
5. Rights: original package MIT; third-party code/media/fonts/music/services retain
separate terms. Remotion production eligibility and unknown asset rights require resolution.
6. Engine isolation: seconds, zero-based frames, caption milliseconds and upstream
inclusive one-based frame tables must be explicitly translated; no mixed CLI/API namespaces.
7. Capability: probe tools and create the best authorized concrete artifact available;
missing tools/rights/facts are reported, and no purchased service is presumed.
8. Honesty: static success, preview, encode, sampled frames, full-motion viewing and
listening establish separate states; cinematic/language/audio/diffusion coverage has limits.
## References
| Resource | Load when |
|---|---|
| [Prompt patterns](references/prompt-patterns.md) | Normalize input and write a coding brief |
| [Director playbook](references/director-playbook.md) | Fill shot/style/timeline/asset fields |
| [Video types](references/video-types.md) | Choose one of seven authored guides |
| [Engine adapters](references/engine-adapters.md) | Probe, route and implement one engine |
| [Quality rubric](references/quality-rubric.md) | Inspect/revise actual artifacts |
| [Operation cards](references/sop_models.md) | Match exact admitted scopes before acting |
| [Source index](references/sources.md) | Verify author/version/rights |
| [Case studies](references/case-studies.md) | Consult provenance and source-local examples |
| [Research notes](references/research_notes.md) | Understand coverage and verification limits |
| [Provenance](references/provenance.json) | Trace rules, nulls, bindings and aliases |
| [Package guide](README.md) | Install locally, browse examples/tests and run checks |
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-changelog-updatemorluto/reaPrepare or rewrite REA release changelogs and GitHub release notes from pinned Git history, with verified contributor thanks and Release Please synchronization. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| academy-guideanthropics/skillsStop and check this skill before finishing any reply to a question about how to use Claude or a Claude product — it recommends matching courses, tutorials, and use cases from Claude Academy (academy.claude.com), Anthropic's learning hub. Trigger on: "how do I", "how can I", "getting started with", "what can Claude do", "teach me", "learn to use"; questions about artifacts, projects, skills, plugins, connectors, MCP; requests about rolling Claude out to a team, class, or organization; and any ask for training materials, onboarding content, or learning resources. Use it when the user is learning how to use a feature or product — not when they are mid-task and just want the task done. This skill composes with other skills: after consulting product documentation to answer how a Claude feature works, also check here for a matching course or tutorial — a docs-grounded answer and an Academy recommendation belong together. Only recommend on a strong match; never invent Academy content. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| claude-apianthropics/skillsReference for the Claude API / Anthropic SDK — model ids, pricing, params, streaming, tool use, MCP, agents, caching, token counting, model migration. TRIGGER — read BEFORE opening the target file; don't skip because it "looks like a one-liner" — whenever: the prompt names Claude/Anthropic in any form (Claude, Anthropic, Fable, Opus, Sonnet, Haiku, `anthropic`, `@anthropic-ai`, `claude-*`, `us.anthropic.*`, `[1m]`); the user asks about an LLM (pricing/model choice/limits/caching) — never answer from memory; OR the task is LLM-shaped with provider unstated (agent/MCP/tool-definition/multi-agent/RAG/LLM-judge/computer-use; generate/summarize/extract/classify/rewrite/converse over NL; debugging refusals/cutoffs/streaming/tool-calls/tokens). SKIP only when another provider is being worked on (overrides all triggers): OpenAI/GPT/Gemini/Llama/Mistral/Cohere/Ollama named in the query; OR `grep -rE 'openai|langchain_openai|google.generativeai|genai|mistralai|cohere|ollama'` over the project hits (run this grep FIRST if no provider named — don't Read the file). | Stars: 198k | Updated | Claude CodeCodex | Unverified | Install |
| doc-coauthoringanthropics/skillsGuide users through a structured workflow for co-authoring documentation. Use when user wants to write documentation, proposals, technical specs, decision docs, or similar structured content. This workflow helps users efficiently transfer context, refine content through iteration, and verify the doc works for readers. Trigger when user mentions writing docs, creating proposals, drafting specs, or similar documentation tasks. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |