本文へ移動
cccskills
無料GitHub で公開

agentsope-ai-video-director

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。

インストール方法を見る

含まれるファイル(27)

  • SKILL.md13.9 KB
  • examples/data-explainer.md10.4 KB
  • examples/hyperframes-prompt.md5.3 KB
  • examples/narrated-docs.md8.8 KB
  • examples/product-launch.md9.7 KB
  • examples/remotion-prompt.md6.1 KB
  • examples/sentence-to-director-plan.md3.9 KB
  • LICENSE1.5 KB
  • README.md7.5 KB
  • references/case-studies.md88.4 KB
  • references/director-playbook.md37.2 KB
  • references/engine-adapters.md85.2 KB
  • references/prompt-patterns.md14.2 KB
  • references/provenance.json669.6 KB
  • references/quality-rubric.md14.8 KB
  • references/research_notes.md7.4 KB
  • references/sop_models.md111.4 KB
  • references/sources.md72.1 KB
  • references/video-types.md23.6 KB
  • skill.json2.3 KB
  • tests/engine-routing.json3.9 KB
  • tests/manual-review.md11.6 KB
  • tests/missing-and-errors.json3.6 KB
  • tests/quality-checklist.md3.0 KB
  • tests/sentence-inputs.json2.8 KB
  • tests/validate_outputs.py30.9 KB
  • tests/validate_package.py13.5 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

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 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. 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 and director fields. 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. 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 or Remotion example. 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 and checklist. 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 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.

ModelUse within its card’s scopePrimary action
H1 / P02Optional HyperFrames prompt controlChoose delegated aesthetics or explicit layout/timing
H2 / P05–P06Optional long/structured explainersWord-led question or scene-frame budget
H3 / P09Optional Lemo dense sonification/occlusionStyle-local note/pen adjustments
H4 / P15TikTok auction ad placement onlyMatch its actual overlay template
M1 / P16–P19HyperFrames 0.8.142 GSAP profileSeconds, root/IDs, readiness, ownership
M2 / P21–P24Remotion 4.0.534 isolated routesFrame state, export, resource handles
M3 / P25–P26/P31Released Motion Canvas 3.17.2Editor exporter, audio, event metadata
M4 / P28/P40Ready/pending voice; typed unitsProvisional audio or measured clock translation
M5 / P37–P38Intended use and source-derived assetsResolve 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. 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.

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

ModeConcrete deliverableState rule
Director planBrief, shots, style, timeline, assetsPlan available; no render implied
Engine promptIsolated versioned coding-agent promptPrompt available; no code implied
ImplementationOriginal code/assets and execution logState actual files and blocked steps
Rendered draft / QAEncoded artifact and covered checksEncoding and review remain independent
RepairDefect, code/timing change, rerun evidenceOnly affected checked states advance
HandoffPaths, states, commands, rights, gapsUnknown/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

ResourceLoad when
Prompt patternsNormalize input and write a coding brief
Director playbookFill shot/style/timeline/asset fields
Video typesChoose one of seven authored guides
Engine adaptersProbe, route and implement one engine
Quality rubricInspect/revise actual artifacts
Operation cardsMatch exact admitted scopes before acting
Source indexVerify author/version/rights
Case studiesConsult provenance and source-local examples
Research notesUnderstand coverage and verification limits
ProvenanceTrace rules, nulls, bindings and aliases
Package guideInstall locally, browse examples/tests and run checks

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Cross-framework enhancement overlay for choosing a multi-agent topology BEFORE writing any agent. A binary-question rubric — is single-agent + tools enough? do agents need to know about each other? does the output need one voice? — maps the answer to single-agent / supervisor / swarm / sequential / hierarchical. Activates when a coder agent is tempted to "split the work into roles" or reaches for a multi-agent framework. Encodes the *selection rubric* that the per-framework skills assume but never surface. Search keywords: when to use multi-agent, single vs multi agent, do I need multiple agents, supervisor vs swarm, multi-agent vs single agent, agent team design.

日本語の概要は準備中です。原文の説明を表示しています。

agentsope/SkillAlchemy4432026年10月9日 更新

SOP for terminal-based, git-native AI pair programming with Aider (git work-tree + tree-sitter repo-map + edit-format + human-in-loop REPL). Use when editing code in an existing git repo via an LLM, when you need to converge a change to 2-5 files, pick an edit format that fits the model, run architect+editor mode, or wire an auto-test loop.

日本語の概要は準備中です。原文の説明を表示しています。

agentsope/SkillAlchemy4432026年10月9日 更新

Screens biomedical / life-science papers for signs of data fabrication, image manipulation, and statistical anomalies, using the detection techniques distilled from the field's canonical exposure platforms (PubPeer, Data Colada, Science Integrity Digest, For Better Science) and tools (ImageTwin/Proofig, statcheck, GRIM/GRIMMER, Problematic Paper Screener, Seek & Blastn). Use when asked to check a paper/figure for image duplication, blot splicing, impossible statistics, paper-mill or tortured-phrase signals, research integrity, or "is this data faked"; or when a user shares a figure, Western blot, supplementary dataset, or DOI and asks whether it looks manipulated. Reports observable anomalies as questions for clarification — it never accuses anyone of fraud.

日本語の概要は準備中です。原文の説明を表示しています。

agentsope/SkillAlchemy4432026年10月9日 更新

Universal discipline for any LM-driven loop — agent retries, plan-act-observe, multi-agent handoffs, optimiser passes, test-fix cycles. Encodes the one rule every framework documents quietly and every team relearns expensively: the LM in the loop is NEVER a reliable terminator. Termination must be provided by an explicit counter + exit predicate + stagnation signal + escalation path that live OUTSIDE the LM's control. This is a tool- level, framework-agnostic skill. It maps onto LangGraph (recursion_limit + state counter + interrupt), CrewAI (max_iter + max_rpm + human_input), Claude / OpenAI SDKs (max_iterations + tool_use_budget), DSPy (declared evaluation budget), Aider (REPL + explicit retry cap), and AutoGen (max_consecutive_auto_reply). Search keywords: infinite loop, recursion limit, recursion_limit, GraphRecursionError, max iterations, max_iter, agent stuck, agent won't stop, runaway agent, ReAct loop not terminating, agent repeating itself.

日本語の概要は準備中です。原文の説明を表示しています。

agentsope/SkillAlchemy4432026年10月9日 更新

Decision rubric for when an LM agent should write-and-run code (Program-of-Thought / code interpreter) versus reason in natural language: classify each step as deterministic- computable (emit + execute code, feed the result back) vs judgment (stay in prose). Use when designing or debugging an agent step that does arithmetic/parsing/data transforms, when prose reasoning hallucinates a computation (under-coding), or when a sandbox round- trip is wasted on a judgment task (over-coding). Search keywords: code interpreter, agent does math wrong, calculator hallucination, when to run code vs reason, program of thought, PoT, tool vs reasoning.

日本語の概要は準備中です。原文の説明を表示しています。

agentsope/SkillAlchemy4432026年10月9日 更新

Coder-agent working-file budget discipline: keep the editable working set (files you /add into writable context) under ~25k tokens, separate "read" from "edit", delegate breadth to a read-only repo-map, and drop files once edited. Use when an LLM coder-agent edits multiple files, when the working set must stay focused, or when the model starts editing the wrong file / missing targets because too much context dilutes attention. Search keywords: context window full, agent edits wrong file, too much context, /add /drop files, working file budget, context dilution, lost in the middle.

日本語の概要は準備中です。原文の説明を表示しています。

agentsope/SkillAlchemy4432026年10月9日 更新

agentsope のスキルをすべて見る

このスキルの問題を報告する