short-form-edit

1How It Works & Mechanism

The sole first-priority gate is 'does the first 3 seconds convince the viewer that they must watch until the end?'; visual effects, footage volume, and production polish all rank behind it, and remain mere editorial judgment until audience data is obtained.

Authorization is a persistent state: a single instruction to 'make the final cut and iterate' authorizes local editing and rendering, requiring additional inquiries only for missing approvals rather than re-asking on every draft; however, publishing or contacting third parties always requires explicit authorization.

Create three markedly different opening drafts using the same raw footage, compare them across topic clarity, curiosity, muted intelligibility, and follow-through intent, and then record the rationale for the final selection.

Open loops are structured: every honest audience question must have an established setup, injection timestamp, reminder or partial answer, resolution timestamp, and the spoken transcript evidence that justifies the resolution.

Graphics and captions are bound to word-level timestamps: captions and impact events must land within ±2 frames of the word onset, allowing up to 0.15s of intentional visual pre-roll.

Brand assets must be authentic: redrawing from memory, approximating, or AI-generating another party's logo is prohibited; if verifiable assets cannot be retrieved, substitute them with plain-text brand names.

Run two local validators (plan structure + footage ledger) before delivery, explicitly acknowledging that validators cannot certify taste, motion aesthetics, or auditory appeal.

Strict mixing specifications apply to audio: normalize voice first, duck the music bed to approximately 14–20 dB below the vocal track before adjusting based on actual tracks, targeting a final output of approximately -16 LUFS and ≤ -1 dBTP.

B-roll accounting and reuse discipline: use the same source scene only once per video (changing crop, speed, or color grade does not count as a new scene), logging entries in the ledger to run validation.

2Core Capabilities

3External Dependencies

4Recommended Governance & Evidence

5Second-Pass Verification

• [ok] 两个自带脚本是否真的不联网、不读凭证（决定档位下限） — validate-plan.mjs 只有 fs/path import 与纯函数导出；validate-footage.mjs 为 fs/path/crypto。全目录 grep 无 process.env、无 fetch/axios、无 http(s):// 调用点（文档中的 URL 仅示例命令与文档引用）。
• [ok] 转写/生成服务是否真由本 skill 声明（橙档依据） — SKILL.md 原文 'Transcribe with the student's chosen provider; ElevenLabs Scribe is Nate's default.' 与 'Kie.ai is Nate's optional provider for generated moving inserts and images.'，两处都要求先看 docs/TOOLS-AND-API-KEYS.md 的凭证与成本。即：外发是**流程指示**，不是随包脚本行为——报告中两者分开陈述。
• [discrepancy] 「禁用 Pointer Lock/pointer capture」是安全收紧还是绕过防护 — ［判定：判定为收紧，不影响档位］SKILL.md 'Browser automation is headless with Pointer Lock and pointer capture disabled in every context.' 出现在「跑 lint 与看实时预览」段，目的是避免无头环境下指针锁抢占；它是关掉被测页面夺权，而非绕过平台自动化侦测（无 AutomationControlled、无真实键击替身、无 remote-allow-origins）。不构成红档特征。
• [ok] 校验器能否替代「品味/动效/听感」复核 — SKILL.md 明写 'The plan validator catches structural defects but cannot certify taste, motion, or the audio listening pass'，并要求另做不看屏的娱乐性复核（ENTERTAINMENT-REVIEW.md）且如实披露缺失的复核模态。脚本与人类判断的分界自洽。
• [unlocatable] 「前 3 秒」这类判断是否有可核的客观阈值 — 门禁条款要求记录观众收益、未解疑问、第 3 秒可见证据与确切回报场景，但「具体到第几帧算成立」没有客观阈值；quality-gates.md 只给失败样例（如 attention grabbed with no reason to continue）。本次只读侦查无法代作者给出可复现判定，故如实标为无法定位为客观标准。
• [ok] SKILL.md 指名的仓库级文档是否真的存在（避免文档断链） — docs/TOOLS-AND-API-KEYS.md、MOTION_PHILOSOPHY.md 与 references/{quality-gates,plan-schema,curiosity-and-entertainment,reference-analysis,premium-motion-and-footage}.md 在仓库内均可定位（非虚构引用）。

6Verdict & Scenarios