Home / All Skills / Content & Media Creation / video-intake-and-storyboard
English (EN)简体中文 (ZH)日本語 (JA)
Content & Media Creation · saranambiar/hyperframes-video-agent-skills

video-intake-and-storyboard

Convert rough video prompts, scripts, reference images, and edit notes into decision-complete scene specs and storyboards for agent-led video production. Use when a user describes a scene, gives timestamped narration, asks to preserve previous scenes, references a concept image, or requests a multi-scene video plan.

L4 High-Privilege · Graded auth & full auditV1 Commit-PinnedExecute CommandWrite FileDual-pass Audit
Author saranambiarGitHub saranambiar/hyperframes-video-agent-skills ↗Stars 48License Apache-2.0（仓库根 LICENSE 为 Apache License Version 2.0；GitHub API license.spdx_id = Apache-2.0）Commit f6b043d9fd
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

The goal is hardcoded as 'producing a spec that another agent can execute without guessing'—the evaluation metric is executability, not document completeness.

skills/video-intake-and-storyboard/SKILL.md
Use this skill before building scenes. The goal is to transform conversational direction into a spec that another agent can execute without guessing.
What is happening here: Defining the intake acceptance criteria as 'downstream does not need to guess'; the related end-to-end-video-playbook also treats 'converting intent into production artifacts first' as step 2.

Intake is extracted across four passes: hard facts (scene numbers, duration, target frame rate/resolution, narration and phrase timestamps, reference images or end-frame targets, asset locations, explicit do-not-touch boundaries, required transitions with adjacent scenes), style constraints, approval state, and missing decisions.

skills/video-intake-and-storyboard/SKILL.md
1. Extract hard facts.
What is happening here: Splitting casual user input into hard constraints versus preferences; explicit 'do not touch' boundaries are classified as hard facts and consistently treated as constraints going forward.

Questioning threshold: Ask only questions that will change implementation, and prioritize concrete choices (crop or do not crop, subtitle style, transition direction, end-frame priority, music volume behavior) rather than open-ended questions.

skills/video-intake-and-storyboard/SKILL.md
- Ask only if the answer changes the implementation.
What is happening here: Minimizing clarification overhead—open-ended questions force the user to do design work for the agent, whereas concrete options enable immediate decisions.

The scene spec quality threshold consists of ten questions: what happens on the first frame, what happens on the last frame, what changes over time, which narration or audio drives timing, which exact timestamp is most critical, which previous scene must remain untouched, which assets are required, which styles must be reused, which visual references dictate layout, and which proof frames demonstrate success.

skills/video-intake-and-storyboard/SKILL.md
A scene spec is ready to implement when it answers:
What is happening here: Providing determinable criteria for 'when a spec is complete,' turning 'sufficient readiness' into a checklist rather than a subjective feeling.

Timestamps are treated as production constraints with semantic mappings ('At 5s' means must be visible at that exact moment; 'Start around 3s' means may start slightly earlier to be visible by then; 'Land by 10s' means no later than; 'Hold for 3s' means no major layout changes during that period; 'Do not appear before' means preset initial state and reveal only when reached).

skills/video-intake-and-storyboard/SKILL.md
"Start around 3s" means animation may begin slightly earlier if it must be visible by 3s.
What is happening here: Converting natural language time expressions into actionable constraints; the related scene-continuity also grants final arbitration over default timing to narration pacing.

Timestamps must explicitly indicate whether they are 'scene-local' or 'video-global'—this is the common root cause of multiple cross-scene failures (subtitles, scene duration, proof frames).

skills/video-intake-and-storyboard/SKILL.md
Always record whether timestamps are local to a scene or global to the full video.
What is happening here: Eliminating an entire class of ambiguity in a single sentence; downstream audio-sync-assembly scene start maps and render-qa target timestamps rely on this convention.

Approval boundary phrases are explicitly defined as constraints rather than suggestions ('Don't change Scene 01', 'Treat as approved', 'Do not regenerate from scratch', 'Keep style, fonts, colors, and components', 'Only update this transition', 'Attached image is for reference only').

skills/video-intake-and-storyboard/SKILL.md
These are constraints, not suggestions.
What is happening here: Turning verbal user permissions into enforceable constraints, aligned with AGENTS.md's 'Never mutate approved scenes unless the user explicitly asks'.

Separation of content and patterns: Notes are split into 'project-specific content' (names, logos, narration, product UI, brand copy, private filenames) and 'reusable patterns' (camera wipes, chart drawing, subtitle preview, crop safety, music ducking, end-frame proof), with only the latter entering future skills or examples.

skills/video-intake-and-storyboard/SKILL.md
Only reusable patterns belong in future skills or examples.
What is happening here: Setting rules for 'distilling skills from real projects'—preventing proprietary assets and brand info from leaking into public repositories (matching README contribution guidelines and AGENTS.md repository boundaries).

2Core Capabilities

01Four-pass intake scanning (hard facts / style constraints / approval state / missing decisions) and the 'only ask questions that alter implementation' threshold
02Scene spec rules: one templates/scene-spec.md per scene; beats written as timestamped visual actions rather than vague motion vibes; explicitly stating what must appear and must not appear early at exact timestamps; including first/last frame requirements and proof timestamps for downstream QA
03Storyboard rules: use templates/storyboard.md for multi-scene workflows; mark approved scenes as locked; annotate carryover dependencies; record which audio controls scene duration and which subtitle or script file is authoritative
04Reference image handling: distinguishing must-match items (composition, colors, sizing, text, positioning) from permissible differences (exact replication, live data, logo graphics, camera cropping, motion paths); reasoning backwards from the last frame; planning end-frame proofs and next-scene carryover proofs
05Approval boundary capture (6 phrase categories) and 'iteration-reducing prompt patterns' (end-frame targets, exact time critical elements appear, visible content before and after that moment, which existing scenes cannot be touched, whether styling can change, whether audio is reference-only or must be overlaid)
06Storyboard failure mode checklist (scene duration mismatches audio duration, scene depends on uncreated previous scene's end frame, reference image conflicts with written timing, user subsequently asks to ignore reference, modifying the final scene cascades into changes to prior subtitles or audio)
07Clarification question checklist (is audio final or reference-only, is the end frame an exact or directional match, can screen recordings reveal the canvas, do silent sections need subtitles, does music swell during pauses) and handoff format (filled spec and storyboard, asset manifest, locked scenes list, open issues, recommended implementation order)

3External Dependencies

KindDependency / Endpoint
package仓库根模板 templates/scene-spec.md（每场景一份）与 templates/storyboard.md（多场景结构）——本件点名使用；位于仓库根，不在 skill 目录内
package输入素材：用户提示、带时间戳的旁白或脚本、参考图、编辑注记与项目内既有场景/渲染（本地读取，无外部服务）

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/: Execution of local scripts or commands: (No scripts exist within this skill directory, and the documentation cites no commands) This skill solely reads and writes documentation: filling specs and storyboards is performed by the host writing files within the project
  • 本地工作目录: File I/O and artifact output: Per-scene scene spec files (filled versions of templates/scene-spec.md)
Audit Notes: The skill directory contains only SKILL.md (pure instructions, no scripts, no endpoints, no credentials); its behavioral scope is local reading (project assets, reference images, existing scenes) and local disk writing (filling specs, storyboards, and manifests), which aligns with the Blue category of 'local read/write only'. **No assets are sent to external services**: full scanning of this skill directory reveals no http(s):// endpoints, no fetch/requests, no token/env reading, and no CLI commands. Its explicit rule separating 'proprietary content from reusable patterns' provides positive reinforcement for repository hygiene (preventing proprietary assets and branding from entering public examples). Evaluated under the same criteria as the nine sibling skills: the directory contains only SKILL.md, behaviors are restricted to local read/write and local CLI, with zero endpoints, zero credentials, and zero outbound asset transfers.

5Second-Pass Verification

  • [ok] 「无外部依赖、无端点无凭证」是否成立 — 目录内正则扫描零命中（无 http(s)://、无 fetch、无 token/env、无 CLI 名）；本件与 motion-design-systems、scene-continuity-and-transitions 同属族内「纯判据、零 CLI」的一组。
  • [ok] 两个模板能否承载 SKILL.md 的质量门槛 — templates/scene-spec.md 含 Beat Timeline（Time、Narration or cue、Visual action、Required proof）与 Continuity（First frame must match、Last frame must be reusable by、Carryover frame files to update）与 QA Checklist（末帧与边界与字幕与音频证明时间戳、已知风险、验收标准），逐项对应十问；templates/storyboard.md 的 Scene Table 含 Status、Starts from、Ends on、Do not touch，并有 Global Rules 的 Approved scenes，对应 locked 与 carryover 依赖要求。
  • [ok] 时间戳语义是否与族内时序真源冲突 — 本件把时间戳当约束并允许为在 3s 可见而略早开始，最终裁决交给音频（- Track which audio clip controls each scene duration.）；audio-sync-assembly 以最新批准的旁白音频为时序真源，scene-continuity 声明 narration timing wins when explicit.——三处一致。
  • [ok] 「不要问可查明的」是否与「只问会改变实现的」自相矛盾 — 两条是同一门槛的两面：先声明只问会改变实现的，再列 5 类可从文件查明的问题（时长、场景名、帧率、含标签的源文件、是否有音轨），实际规则是「能读盘就不问，读了还不确定才问」。无矛盾。
  • [ok] 本件是否覆盖「用户说忽略之前参考图」这类撤回场景 — Storyboard Failure Modes 列出 A user saying "ignore this reference" after it was already used.，属对撤回的提示；但未给具体回滚程序（只提示风险），实际回滚需依赖 render-qa 的外科流程——属分工而非缺口。
  • [discrepancy] 「内容与模式分离」是否真能防止私有素材出仓 — 规则只要求把项目专有内容与可复用模式分开，并说私有文件名等属专有内容；但 SKILL.md 未给可执行的检查（例如私有资产路径清单或提交前审查），实际执行依赖 agent 自觉与仓库 AGENTS.md 的 Keep private project assets out of this repository.。属规则层约束，非机制。

6Verdict & Scenarios

  • Sets the intake acceptance criteria to 'downstream agent does not need to guess' with a 10-question threshold, making 'readiness' objectively determinable
  • Timestamp semantic mapping plus local vs. global labeling eliminates the primary source of ambiguity across multi-scene, multi-track setups in a single rule
  • Defines approval boundary phrasing as hard constraints, directly backing all downstream stop conditions for 'do not mutate locked scenes'
  • Questioning discipline strictly prioritizes concrete options (crop or not, subtitle style, transition direction, end-frame priority, music ducking behavior), minimizing clarification costs
  • Separation of content and patterns simultaneously ensures spec quality and repository hygiene (keeping proprietary assets and brand info out of public examples)
    • Best for: Best suited for: The project kick-off phase—when the user provides only rough ideas, timestamped narration, reference images, or edit notes, and needs them transformed into an actionable multi-scene plan; also ideal when adding new scenes, preserving existing scenes, or requiring a spec and storyboard that 'another agent can directly execute'.
    Not for: Not suited for: (1) Situations where a complete, approved spec already exists and direct implementation begins (use hyperframes-scene-builder instead); (2) Minor finishing tweaks (use render-qa-and-surgical-changes); (3) Scenarios requiring actual video/audio output or acceptance evidence (this skill has no commands or scripts and only outputs documentation); (4) Projects requiring formal compliance reviews or proprietary asset redaction pipelines (the separation rule is a constraint, not an automated mechanism).
    Install Agent-Ready
    ① Verified Mirror 2026-09-28
    Option A · Direct DownloadDownload video-intake-and-storyboard.tar.gz
    sha256: f997b1af0e9338b6…
    Option B · Copy Guide for Agent
    install.json
    Agents automatically download and verify mirrors from this JSON guide.
    ② Upstream GitHub
    GitHub Repo ↗
    Pinned to commit f6b043d9fd.