Content & Media Creation · saranambiar/hyperframes-video-agent-skills
end-to-end-video-playbook
End-to-end video production workflow for coding agents using HyperFrames, FFmpeg, captions, audio sync, music beds, and render QA. Use when a user asks for a complete video, multi-scene edit, launch video, product/demo video, UGC-style edit, enterprise explainer, or final assembled MP4 from prompt/script/assets.
L4 High-Privilege · Graded auth & full auditV1 Commit-PinnedExecute CommandWrite FileDual-pass Audit
Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.
1How It Works & Mechanism
This skill serves as the master orchestration entry point for the entire family: moving from a rough request to a verified final render via a five-step main workflow (grounding -> converting to production artifacts -> routing to specialized skills -> layered building -> verifying before reporting).
skills/end-to-end-video-playbook/SKILL.md
Use this skill to coordinate a complete video project from rough request to verified final render.
What is happening here: Structuring work into the sequence of 'inspect what already exists on disk first', then 'produce executable artifacts', and finally 'dispatch to specialized skills', preventing coding scenes right from the start.
The very first step requires identifying 'locked scenes' and user boundaries (e.g., do not touch Scene 01), and mandates using a storyboard template to carry the multi-scene structure.
skills/end-to-end-video-playbook/SKILL.md
Identify locked scenes and user boundaries such as "do not modify Scene 01."
What is happening here: Treating 'approval status' as a first-class input; all stopping conditions for downstream surgical skills are built on this 'locked' concept.
Convert intent into artifacts: one scene-spec per scene, strictly enforcing the separation of 'project-specific content' from 'reusable production patterns' (preventing private assets/branding from leaking into reusable examples).
skills/end-to-end-video-playbook/SKILL.md
- Separate user-specific content from reusable production patterns.
What is happening here: Corresponds to repository AGENTS.md rules ('Keep private project assets out of this repository') and README contribution principles—the workflow itself enforces the boundary discipline between 'reusable' and 'private'.
Includes a built-in routing table that directly maps user phrasing to 8 specialized skills (e.g., rough idea only -> intake; reference final frame provided -> intake -> motion-design-systems -> scene-builder; 'everything perfect except X' -> render-qa).
skills/end-to-end-video-playbook/SKILL.md
- User says "everything is perfect except": use `render-qa-and-surgical-changes`.
What is happening here: This is the most practical part of this skill—turning 'which skill should be loaded' into a table lookup rather than letting the agent guess.
The layered build order is a hard constraint: draft scene renders first -> music-free narration master first -> captions -> music bed last; rules explicitly forbid patching an old final render when source scenes and assembly scripts exist.
skills/end-to-end-video-playbook/SKILL.md
- Assemble narration-only master before final music.
What is happening here: Using ordering to proactively eliminate drift like 'editing captions/music against a mix that already contains music', aligning with the layered prerequisites of audio-sync-assembly and captions-and-music-bed.
Completion criteria is defined by a Quality Gate checklist: final duration/resolution/FPS known, audio tracks present, captions present (if required), scene boundary proof exists, user-specified timestamp proofs exist, music tail behavior verified, and subjective previews requested by the user are shown or explicitly skipped.
skills/end-to-end-video-playbook/SKILL.md
- Exact user-requested timestamp proofs exist.
What is happening here: Defining 'done' as a set of observable proofs, requiring a declaration at the end of 'what is verified vs. what remains subjective'.
Questioning discipline: Only ask questions that will alter the implementation, and explicitly forbid asking questions that can be discovered from local files (clip duration, scene names, current FPS, which source file contains a specific tag, presence of audio tracks).
skills/end-to-end-video-playbook/SKILL.md
Do not ask questions discoverable from files:
What is happening here: Codifying 'read the disk before asking' into a rule, effectively requiring the agent to eliminate uncertainty using tools; the sibling video-intake-and-storyboard reiterates this exact rule.
2Core Capabilities
01End-to-end orchestration and routing: 5-step main workflow + 12-step implementation order (lock scenes -> prepare spec -> build source -> render drafts -> extract proof frames -> music-free assembly -> caption alignment -> burn captions -> music bed -> audio mix -> export final cut -> proof sheet and preview clips)
02Artifact delivery checklist: storyboard or timeline manifest, scene renders, final assembly script or command plan, caption files, final MP4, proof frames/contact sheets, risk segment preview clips
03Decision checklist (10 items, must be logged before building): new video/add scenes/finishing touches, whether existing scenes are approved, whether old renders are true sources or previews only, captions needed or not, music add/replace/duck/keep, whether to crop product demos, whether exact speech timing is required, presence of silent hooks/pauses/tail holds, final FPS set to 24/30/48/60 or source rate, whether the user expects proof before rendering
04User-phrasing-based routing table (rough idea / reference final frame / scene must start from another scene / raw screen recording / captions late or too big / narration changed / everything perfect except)
05Four production rules: do not patch old final renders when sources exist, do not slow down globally for a small pause, prioritize hold frames for silence and pacing, re-render only affected scenes for finishing changes and check for duplicate carryover frames
06Question checklist (only the 6 that change implementation) and 5 'do not ask' items discoverable from files
07Handoff format: final render path, key scenes and render paths, commands or scripts used, proof frames/preview paths, tests conducted, unverified items
3External Dependencies
| Kind | Dependency / Endpoint |
|---|---|
| cli | HyperFrames（场景合成与渲染引擎；本件在 description 与「构建场景」路由中点到，具体命令由 hyperframes-scene-builder 给出） |
| cli | FFmpeg / ffprobe（装配与探针；文中以 media probe 表述，具体命令在 audio-sync-assembly 与 render-qa-and-surgical-changes 中） |
| package | 兄弟 skill（同仓 8 件：video-intake-and-storyboard、hyperframes-scene-builder、motion-design-systems、scene-continuity-and-transitions、product-demo-integration、audio-sync-assembly、captions-and-music-bed、render-qa-and-surgical-changes）——被本 skill 按名点名调用，属宿主/仓库层协作，不改变本件档位 |
| package | 模板 templates/storyboard.md 与 templates/scene-spec.md（本件要求用它们承载多场景结构与逐场景 spec） |
| package | 素材输入：项目内既有资产、渲染、组合、脚本、音频、字幕与既往批准（本地读取，非外部服务） |
4Recommended Governance & Evidence
Suggested Controls for Safe Operation
For Agent Runtime
Graded/per-use authorization, sandbox recommended
For Administrators
Approval workflow, rollback support, full audit
Evidence & Triggers (File Paths & Excerpts)
- scripts/: Executes local scripts or commands: (No built-in scripts in this skill directory) Executed by host according to routing: ffprobe/ffmpeg probing and assembly, HyperFrames rendering, captioning, and audio mixing commands
- 本地工作目录: File I/O and artifact output: storyboard, timeline manifest, per-scene scene-spec (production artifacts)
Audit Notes: This skill directory only contains SKILL.md (pure workflow/routing documentation, no scripts, endpoints, or credentials); behavioral scope is reading project-internal files, writing production artifacts (spec/storyboard/renders/captions/proof frames), and invoking local CLIs and sibling skills per the routing table, conforming to the Blue Tier ('local I/O / local tools only'). **No assets are sent to external services**: Full scan of this skill directory reveals no http(s):// endpoints, no fetch/requests, and no token/env reading. The 8 mentioned sibling skills share the same repository and tier (pure prompts + local tools); per paradigm, tiers are not transitively escalated, and collaboration is noted in reason and risks. All nine sibling skills share identical classification criteria: directory contains only SKILL.md (+ references), behavioral profile is local file I/O and local CLI, with zero endpoints, credentials, or outbound asset transmission.
5Second-Pass Verification
- [ok] 路由表点名的 skill 是否都存在（避免指向不存在的 skill） — 本仓 skills/ 下确为 9 个目录：audio-sync-assembly、captions-and-music-bed、end-to-end-video-playbook、hyperframes-scene-builder、motion-design-systems、product-demo-integration、render-qa-and-surgical-changes、scene-continuity-and-transitions、video-intake-and-storyboard；路由表点到的 8 个全部存在，含自身构成完整闭环。
- [ok] 是否存在隐藏网络调用（编排层最容易藏外发） — 全目录仅一个 SKILL.md，正则扫描无 http(s):// / fetch / requests / api_key / token / .env / curl；所有外部能力都表现为「本地 CLI 或同仓 skill」，无素材上传与生成服务。
- [discrepancy] Run media probes 是否指得出具体命令 — 本件只说 Run media probes on final outputs.（步骤 5）而未点名 ffprobe 或仓库脚本；具体命令分散在 render-qa-and-surgical-changes（六个脚本清单）与 audio-sync-assembly（ffprobe 与 setpts 模板）。只加载本件的 agent 需自行选择探测方式。
- [ok] Quality Gate 是否可机械判定 — 七项中六项可观察（时长/分辨率/FPS/音轨存在/边界证明帧/点名时间戳证明），一项是流程性（音乐收尾行为已查），一项是主观项并要求显式声明跳过——判据明确，不靠感觉。
- [ok] 5 步主流程与 12 步实施顺序是否冲突 — 两者是层级关系：5 步是阶段划分（ground/intent/route/build/verify），12 步是「构建 + 装配」段的展开，前 5 步对应 build 前的准备、后 7 步对应 build 与 verify；无相互矛盾的指令。
- [ok] 「不要 patch 旧成片」是否在族内一致 — 本件 Production Rules 的 Never patch an old final MP4 when source scenes and assembly scripts exist. 与 render-qa-and-surgical-changes、workflows/surgical-final-edit.md 的同一规则一致（后者另写 Do not patch an old final MP4 if source scene renders and assembly scripts exist.）。
6Verdict & Scenarios
Best for: Best for: The first stop when a user presents an end-to-end request (complete video, multi-scene edit, launch video, product/demo video, UGC-style edit, enterprise explainer, or producing a final MP4 directly from prompts/scripts/assets); also ideal as a persistent entry document for a multi-skill video repository to dispatch the other 8 skills.
Not for: Not suitable for: 1) Point fixes or minor surgical tweaks (if user says 'everything is perfect except X' -> go directly to render-qa-and-surgical-changes, skipping intake artifacts); 2) Tasks requiring concrete implementation details—this skill contains no code, CLI commands, or styling heuristics, and must be paired with specialized skills; 3) Projects requiring asset licensing or compliance review (quality gates do not cover this); 4) Workflows requiring online collaboration or hosted cloud rendering (assumes local CLI and local project directories).
Install Agent-Ready
① Verified Mirror 2026-09-28
Option A · Direct DownloadDownload end-to-end-video-playbook.tar.gz
sha256:
Option B · Copy Guide for Agent70061b50836680c0…Agents automatically download and verify mirrors from this JSON guide.
② Upstream GitHub
GitHub Repo ↗
Pinned to commit f6b043d9fd.