Content & Media Creation · Agentchengfeng/chengfeng-videocut-skills

videocut-skills

Name: videocut-skills
Author: Agentchengfeng/chengfeng-videocut-skills

Edit raw Chinese talking-head footage: word-by-word transcription, dictionary-based typo correction producing a correction ledger, five-pass scanning for speech stumbles and repetitions, summary and repeated-sentence tables, launching Studio for user review, and retrospective analysis to capture user preferences and vocabulary. Only outputs a verified deletion ledger—does not cut media, generate subtitles, or render storyboard animations. Use when the user requests editing talking-head videos, cleaning up stumbles, preparing baseline talking-head footage, continuing a cut, or when a confirmation card returns action=return_cut_review. Do not use for physical video clipping, exporting final cuts, standalone installation, opening the workbench by itself, or producing finished storyboarded videos.

L4 High-Privilege · Graded auth & full auditV1 Commit-PinnedExecute CommandNetwork AccessWrite FileDual-pass Audit

Author AgentchengfengGitHub Agentchengfeng/chengfeng-videocut-skills ↗Stars 2996License Apache-2.0（仓库根 LICENSE 首行 'Apache License'、次行 'Version 2.0, January 2004'；plugin package.json 亦声明 "license": "Apache-2.0"）Commit b10e85e169

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 entry is a Codex Plugin collection (6 business skills) rather than a single-file skill: the primary entry point is 'cut talking-head' chengfeng-cut, registered alongside subtitles, visuals, export, bug reporting, and update checks, with shared boundaries documented in internal references rather than consuming user entry points.

plugins/chengfeng-videocut/.codex-plugin/plugin.json

Six task-facing chengfeng-videocut Skills: cut, subtitle, visual, export, safe bug reporting, and Marketplace update checks. Shared product boundaries are internal references, not a user-facing Skill.

plugin.json's "skills": "./skills/" points to the 6 SKILL.md files under plugins/chengfeng-videocut/skills/; the main entry point resolves to chengfeng-cut based on defaultPrompt[0] ("Use $chengfeng-videocut:chengfeng-cut to cut out speech stumbles from this video and generate a deletion ledger after review.") and the first row of the plugin README capability boundary table (Cut Talking-Head → chengfeng-cut), where skill.path represents this directory.

The main entry point strictly handles semantic judgment and orchestration: it explicitly produces no media, and its sole artifact is a human-reviewed deletion ledger (Cuts + EDL), with ground truth for time, media, and state residing in the product Runtime.

plugins/chengfeng-videocut/skills/chengfeng-cut/SKILL.md

**不切媒体**：账本改一次是几十毫秒，切一次是一个不可撤销的文件。到本 Skill 结束，

Immediately followed by: 'No new video files are created on disk. The Skill handles semantic reasoning and orchestration; the product Runtime is the sole writer for project, Cuts, and Studio state.'—Media writing is entirely moved out of the skill layer.

All state mutations follow Product CAS: first readback to acquire the revision, then submit candidates with --expected-revision, and immediately perform another readback to verify the stage and three revisions (project, Cuts, and EDL).

plugins/chengfeng-videocut/references/business-workflow-contract.md

使用上一步 readback 的 expected revision 调用 `cuts set`、`artifact put` 或 Product transition；随后立即 readback，确认 Product 返回的新 revision 与预期 stage。

The same contract mandates that upon revision_conflict or readback mismatch, it must 'halt, display diff, and re-propose/re-review without overwriting'—failing fast rather than retrying with force overwrites.

The only valid input for deletion reasoning is the product-supplied 'playback order' view (transcript playback); the agent is strictly forbidden from assembling transcript.json manually, which silently misinterprets missing indices as missing content.

plugins/chengfeng-videocut/skills/chengfeng-cut/SKILL.md

**判断只用 `transcript playback` 的播放顺序，不许自己从 `transcript.json` 拼**

references/semantic-deletion.md records two real-world failure cases (grouping by original paragraph numbers and misidentifying compressed pauses as deleted words), concluding that 'the criteria were correct, but the input lied silently.'

Deletion reasoning is split into five single-criterion scan passes (repetitions / false starts & mid-sentence corrections / stumbles & restarts / English hesitations / filler phrases), each pass focusing on one criterion and recording one type, leaving conflict resolution to aggregation.

plugins/chengfeng-videocut/skills/chengfeng-cut/SKILL.md

把播放顺序完整读五遍，每遍只带一个判据。

The rationale in the same section states: 'Scanning with one criterion at a time yields vastly superior accuracy compared to scanning with six simultaneously; each pass generates small, inspectable outputs.'

Aggregation generates two tables with mandatory self-reading verification: Deletion Summary Table + Repeated Sentences Table. The AI reads the post-deletion text aloud per the tables; if it sounds unnatural, the cut is reverted (strictly binary state, no 'suggested deletion' intermediate state).

plugins/chengfeng-videocut/skills/chengfeng-cut/SKILL.md

这是复核的定义：AI 自己再看一遍，不是拿去问用户；拿不准 = 不列（两态）。

Both tables include a risk column (Low/High); long sentences, full sentences, and branched restarts must be flagged for High-Risk review.

The manual review phase binds UI launch to dual verification of project-level binding and capability gating: the URL's #project/<projectId> must match the readback, and ensure-studio must return ok=true before opening the page.

plugins/chengfeng-videocut/skills/chengfeng-cut/SKILL.md

- 打开前把 `productUrl` 的 `#project/<projectId>` 与 readback 的 `projectId` 严格比对，

The same section binds stage=cut_review_ready and three revisions, explicitly declaring that 'a gate returning ok=true only proves product-side capability and cannot replace project-level binding.'

Post-handover retrospective and crystallization: diff the preserved proposal file against the final ledger, archiving differences into three categories (style preferences / proper noun dictionary / rule gaps).

plugins/chengfeng-videocut/skills/chengfeng-cut/SKILL.md

- 口味差异 → `~/.chengfeng-videocut/cut-preferences.md` 追加（Agent 维护的

Preference files and dictionaries are user-level Markdown files; the Agent only appends or updates its own sections. Proper noun errors go to dictionary.md, and rule gaps are logged as 'pending case law upgrades', only suggested for promotion to general rules upon the third recurrence.

Conclusion grading is enforced as hard gating: API/readback PASS, visual frame PASS, and human listening PASS are mutually non-substitutable; if no human has actually listened, it must be reported as UNVERIFIED.

plugins/chengfeng-videocut/references/business-workflow-contract.md

没有真实人类听音记录时，必须报告 **human listening UNVERIFIED**；不得声称 PASS。若本轮没有真实 UI 路径，visual frame 也必须是 UNVERIFIED，而不是由静态测试代替。

The export SKILL.md further clarifies that 'preview and final export are two distinct render pipelines', forbidding preview screenshots as proof of the final export.

Environment logic is centralized into a single dedicated Skill: step 0 of any business Skill only references the 'readiness check' of check-updates, which yields strictly three states.

plugins/chengfeng-videocut/skills/chengfeng-check-updates/SKILL.md

三步走完，输出三态之一：**就绪 / 需新会话 / 停**。

This Skill states: 'All environment logic is isolated to this Skill... any environment policy updates require editing only this single file'; halt states include anti-accident rules prohibiting ad-hoc review pages or custom players from substituting product interfaces.

Skills updates are split into two transactions—'discovery' and 'installation mutation': the check phase parses candidates inside an isolated, temporary CODEX_HOME using official stable releases, executing zero codex calls and zero writes to the user's HOME.

plugins/chengfeng-videocut/skills/chengfeng-check-updates/SKILL.md

正常检查只允许在临时 `CODEX_HOME` 中用官方 `stable` 发现候选；对用户

Code evidence: check-plugin-update.cjs's stageCandidate() runs `plugin marketplace add … --ref stable` and `plugin marketplace upgrade` with isolatedEnv = { ...process.env, CODEX_HOME: codexHome }; candidates must match four identity points: version + 40-hex B(snapshotRevision) + 40-hex A(contentRevision) + SHA-256.

Activation uses only official Codex commands and supports rollback: once the user confirms the four-factor identity, the script re-stages, validates --expected-*, performs remove → add --ref <B> → upgrade, and re-reads from cache to verify.

installer-manifest.json

"codex plugin marketplace add",

installer-manifest.json's allowedCommands fixes the command whitelist to 7 codex plugin commands; SKILL.md mandates that B must be a 40-hex snapshotRevision, treating stable, branches, tags, or short SHAs as marketplace_ref_unsafe.

2Core Capabilities

01Talking-head project initialization and state readback (project create + workflow get + cuts get, both readbacks must share the same projectId)

02Dictionary-based typo correction gate: two-pass validation through general and user dictionaries to generate a typo correction table; modifies only text without touching timestamps, word counts, or wordId

03Five-pass single-criterion semantic deletion (repetitions / false starts / stumbles / English hesitations / filler phrases), recording wordIds, type, and risk per entry

04Two-table aggregation + self-reading verification (Deletion Summary Table, Repeated Sentences Table), strictly binary states without intermediate options

05Studio manual review orchestration: present tables before opening pages, idempotent ensure-running → open to retrieve product URL → ensure-studio capability gating

06Retrospective archiving: diff proposal vs. final ledger, crystallizing user preferences and proper nouns into three distinct categories

07Subtitle phase: calculate post-cut timing from the ledger, storing word ID lists + display text per screen without storing raw seconds

08Visual phase: bind HTML layers to subtitle cues (--cues), extract frames for visual inspection, detect motion windows via pixel diffs, selecting styles strictly from the registry

09Export phase: cut segments + punch-in/push + subtitles + HTML overlays burned into an mp4 in a single pass, tracking four progress stages: assemble → overlay → compose → verify

10Bug reporting phase: sanitized draft → user confirmation token → gh issue create targeting fixed repository and labels

11Update/Environment phase: readiness checks, Runtime installation and health inspection, rollback-safe Skills activation

3External Dependencies

Kind	Dependency / Endpoint
cli	codex（OpenAI Codex CLI）
cli	git
cli	gh（GitHub CLI）
cli	curl
network	GitHub Releases（Runtime 制品与 install.cjs）
network	官方 Codex Marketplace 源（Git）
network	本地 Studio / Product Runtime（loopback 常驻服务）
cli	node ≥ 20 / ffmpeg ≥ 6 / ffprobe / Google Chrome
api	云端 ASR（火山引擎，凭证由 Runtime 保存）
package	MCP 运行时依赖（plugin 包）

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/: Execute local scripts or commands: node plugins/chengfeng-videocut/scripts/ensure-runtime.cjs --install-if-mis
网络端点: Declare external network access: https://github.com/Agentchengfeng/chengfeng-videocut/releases/download/v0.4
本地工作目录: File read/write and artifact output: ~/.chengfeng-videocut/ (Runtime installation directory written by the installer; cut-preferences.md, dictionary.md)

Audit Notes: Risk rating basis (under the 6-tier model: dependencies on remote packages or triggering credential chains require an Orange rating) — ① Downloads and executes remote installers: ensure-runtime.cjs uses curl to pull install.cjs and portable artifacts from official GitHub Releases, then spawns node to run them, representing a remote code execution attack surface (Mitigations: pinned release tag v0.4.11 + SHA-256 checksums + platform gating + fail-closed design); ② Triggers host credential chains: modifies the user's Marketplace/plugin cache via the official codex CLI (rollback-capable write transaction) and creates public issues via gh CLI; ③ Involves third-party services: cloud ASR requires user Volcano Engine credentials, sending audio and transcribed text outbound; ④ Substantial local disk writes: ~/.chengfeng-videocut, user Codex configurations and caches, and project files (via product CLI). Mitigations are comprehensive: write transactions require user confirmation, CAS readbacks, and state-machine-driven error handling; injection vectors are covered by sanitization and parameterized spawn tests. Note: This rating strictly describes operational behavior and does not reflect author trust.

5Second-Pass Verification

[ok] skill.path 定位（任务表标『待定位』；本 repo 是合集形态） — repo 是 Codex Plugin 安装入口 + 一包 6 个 Skill，plugin 根无 SKILL.md；主入口按 plugin.json defaultPrompt[0] 与 plugin README 能力边界表首行取 chengfeng-cut，故 path=plugins/chengfeng-videocut/skills/chengfeng-cut。其余 5 个 SKILL.md 路径已在 analysis.capabilities 与 internal_assets 中逐条登记。
[ok] skill.commit / audit.skill_commit 与 pin 一致 — 工作副本 HEAD b10e85e1694ac562cb9d0a9496e1a79545d10071；任务表 pin b10e85e1694a；PINS-b2.json 记 {dir: chengfeng-videocut-skills, commit: b10e85e1694a}。
[ok] meta.license 与仓库事实 — 仓库根 LICENSE 为 Apache License Version 2.0；plugin package.json 与 .codex-plugin/plugin.json 均声明 Apache-2.0；GitHub API spdx_id = Apache-2.0，无冲突。
[ok] stars / last_push 交叉核对 — gh api 返回 stargazers_count 2996、pushed_at 2026-09-14T15:41:49Z；工作副本最后一次提交时间为 2026-09-14 23:40:09 +0800（同年同日），一致。
[unlocatable] 被顶替关系（本条目顶替 to-prd/writing-great-skills） — 『为什么由本 skill 顶替』属任务分配层决定，源码内无证据可核；本条仅按 agent-18.md 指派执行。
[discrepancy] 池内旧描述（aws_review：FFmpeg + 火山引擎 ASR + 上传 uguu.se 公共图床）是否适用于本 pin — 全仓 grep 'uguu' 零命中（仅 git pack 二进制噪声），该描述来自旧仓库 Ceeon/videocut-skills 时代；本 pin 的实际外部依赖是官方 GitHub Releases/Marketplace、codex/gh CLI 与 loopback 5190。云端 ASR 仍是火山引擎，但凭证经 Runtime 管理。池内 aws_review 与 pool.gh_url（Ceeon/videocut-skills）均为待更新的历史值。
[ok] README/契约中的『已验收/未验收』自述是否与代码一致 — README 明说本轮是 CLI/portable 安装修复预发行、不推进 stable、常驻服务生命周期尚未完成新版本验收；runtime-requirements.json 的 minimumRuntimeVersion 0.4.10 与 releaseTag v0.4.11 与之相符，未发现把预发行包装成 stable 的表述。
[ok] 『user-invocable 只证明兼容 metadata，不证明 UI 展示』是否真在源码写明 — plugin README 原句：静态元数据、agents/openai.yaml、Plugin 首页 starter prompt 和 CLI 的 $plugin:skill 调用，不能单独证明 Desktop Slash/Plugin 群组已经在界面中展示。

6Verdict & Scenarios

Strict separation between semantic reasoning and media/state mutations; the agent only submits reviewable ledger candidates

Every state mutation is backed by a CAS and readback audit trail, failing fast without overwriting user edits

Honest conclusion grading: separates API/readback, visual frames, and human listening, strictly falling back to UNVERIFIED without human listening

Engineering practices informed by post-mortems, ensuring full traceability of design decisions rather than abstract style guidelines

Updates and installations follow rollback-safe transactions with cryptographic identity verification (40-hex + provenance + SHA-256) and a complete failure state machine

Best for: Ideal for one-stop editing of raw Chinese talking-head video footage: intended for Codex users who want a semi-automated pipeline ('AI identifies stumbles/repetitions/false starts → human reviews via two tables in Studio → retrospective crystallizes preferences') and have installed or are willing to install the Chengfeng VideoCut Runtime. Also serves as an engineering blueprint for skill authors studying 'agent orchestration + product CAS + honest acceptance grading'.

Not for: Not suitable for users seeking one-click fully automated video rendering who refuse manual reviews and Studio interaction (the workflow explicitly halts at multiple points for human approval). Not suitable for non-talking-head content (vlogs without word-by-word transcripts, music videos, non-Chinese audio) or environments lacking the required Runtime, Chrome, or cloud ASR. Not suitable for sensitive video footage, as cloud transcription sends audio to Volcano Engine; also unsuitable for purely offline users requiring zero outbound traffic and zero writes to the user environment.

Install Agent-Ready

① Verified Mirror 2026-09-15

Option A · Direct DownloadDownload videocut-skills.tar.gz

sha256: 34c8b28b38500e4f…

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 b10e85e169.