Video Link Analysis Skill V2.2

Turn a video link into an evidence-backed, multimodal Markdown source analysis—not just a summary.

video-link-analysis is a Source Analyzer for Codex and other compatible coding agents. It converts a video into a verifiable, traceable Draft Markdown source analysis that can be uploaded to a Personal Knowledge System. It guides an agent to inspect speech, captions, chapters, key frames, charts, documents, and screen demonstrations; report evidence coverage and limitations; distinguish source facts, creator views, Skill inferences, and external verification conclusions; and propose optional cognitive candidates that remain explicitly unconfirmed by the user.

It preserves three layers:

original video and source facts
→ Skill-generated analysis
→ user-confirmed personal cognition

The Skill produces only the middle layer. It never marks an AI summary, implication, recommendation, or cognitive candidate as the user's adopted view.

V2 does not attempt wasteful frame-by-frame processing. It selects evidence according to the video's content mode and records what was checked, what was not checked, and how strongly the final account is supported.

What it does

• Classifies videos as transcript-rich, hard-subtitle, visual-evidence, screen-demo, visual-narrative, or insufficient.
• Reads available transcripts, captions, chapters, accessibility text, and key frames.
• Requires visual inspection for charts, documents, demonstrations, comparisons, and visually dependent claims.
• Builds a timestamped multimodal evidence ledger.
• Checks agreement between speech, captions, visible scenes, descriptions, and platform summaries.
• Assigns an A–D evidence grade and records coverage and unverified intervals.
• Reports analysis confidence separately from video coverage.
• Maps each material claim to video evidence, external evidence, a verdict, and claim confidence.
• Uses a reader-first layout: conclusions and practical implications first, evidence mechanics in a final appendix.
• Refuses to claim it watched a video when only a title or platform summary was available.
• Separates content facts, external facts, causal claims, predictions, value judgments, and rhetoric.
• Checks important or time-sensitive claims against primary and authoritative sources.
• Looks for both supporting evidence and counterevidence.
• Evaluates assumptions, reasoning gaps, business implications, risks, and applicability.
• Writes a reusable Markdown note with Frontmatter, citations, limitations, and open questions.
• Uses the stable video-analysis/v1 Front Matter contract with status: draft for new output.
• Produces 0–5 clearly labeled cognitive candidates without creating a Cognitive Card.
• Remains readable as standalone Markdown even when no knowledge system parses the Front Matter.

Install

Current stable version: 2.2.0.

With GitHub CLI 2.90 or later:

gh skill install Soda1412/video-link-analysis-skill video-link-analysis \
  --agent codex \
  --scope user

To install the stable v2.2.0 release explicitly:

gh skill install Soda1412/video-link-analysis-skill video-link-analysis \
  --agent codex \
  --scope user \
  --pin v2.2.0

Manual installation for Codex:

git clone https://github.com/Soda1412/video-link-analysis-skill.git
mkdir -p ~/.codex/skills
cp -R video-link-analysis-skill/skills/video-link-analysis ~/.codex/skills/

Confirm the installed version:

cat ~/.codex/skills/video-link-analysis/VERSION

If an older copy already exists, move it to a backup location before installing. Do not copy the new folder inside the existing skill directory.

Usage

Invoke the skill explicitly:

Use $video-link-analysis to analyze this video, fact-check its main claims,
evaluate the reasoning, and save the result as Markdown:
https://example.com/video/123

The description also supports natural-language requests such as:

Analyze this video and tell me what the creator got right or wrong.
Save the analysis as a Markdown research note.

Expected output

The generated document follows a stable reader-first structure:

1. analysis boundary;
2. one-sentence summary;
3. content summary;
4. key views;
5. source facts, creator views, Skill inferences, and external verification;
6. evidence and limitations;
7. implications;
8. optional cognitive candidates, labeled AI-generated and not user-confirmed;
9. bounded suggested actions;
10. open questions;
11. content overview;
12. detailed claim-evidence analysis;
13. source and method appendix.

The included template is adaptive, but the analysis boundary, summary, key views, evidence limits, cognitive-candidate boundary, open questions, and method appendix always remain. Optional detail can be removed when it would only repeat the summary.

Front Matter contract

New output uses schema_version: "video-analysis/v1", document_type: video_analysis, source_type: video, and status: draft. It preserves both the submitted and canonical URLs, source metadata, language, evidence grade, analysis confidence, coverage, visual mode, unresolved intervals, topics, explicit project relationships, and the Skill generator version. related_projects remains an empty array unless a relationship is supported.

evidence_grade evaluates coverage of the original video. analysis_confidence evaluates reliability of the resulting analysis. External sources can improve claim verification but cannot fill unobserved video segments or raise the coverage grade.

Quality and safety boundaries

• A platform-generated summary is not treated as a complete transcript.
• A displayed number is treated as a visible claim, not proof that the number is true.
• A mockup, slide, marketing animation, or success screen is not treated as proof of a deployed working system.
• Evidence grades describe content coverage, not creator truthfulness or analysis quality.
• Comments are audience-reaction samples, not factual evidence.
• The workflow is read-only by default: it does not like, follow, comment, repost, or modify accounts.
• It does not bypass authentication, CAPTCHA, paywalls, regional restrictions, or platform protections.
• It does not download or redistribute copyrighted videos.
• It does not turn an external creator's opinion into a user or project decision.
• It does not automatically create or confirm Knowledge Cards, Cognitive Cards, Captures, or user cognition.
• It does not call or modify a Personal Knowledge System; handoff remains Markdown output followed by user upload.
• It does not execute suggested actions, publish analyses, or bulk-scrape private accounts.
• High-stakes medical, legal, and financial claims require additional caution and authoritative sources.

Requirements

The skill is tool-agnostic, but the host agent needs enough capability to access the supplied source. Depending on the video platform, that may include:

• a platform connector or official API;
• a supported browser with the user's existing session;
• a transcript, subtitle, audio, or video file supplied by the user;
• web search for current fact-checking;
• filesystem access to save the Markdown document.

If none of these provide adequate evidence, the skill requires the agent to state the limitation instead of fabricating an analysis.

Evidence grades

Grade D output must not imply that the full video was watched.

Analysis confidence is a separate document-level judgment. It uses high, medium-high, medium, medium-low, or low to describe whether the final evaluation remains reliable after considering source quality, claim-level fact checking, cross-modal conflicts, and unresolved gaps. Existing V2 documents without this field remain valid; newly generated complete analyses should include it and a claim-evidence matrix.

Validate V2.2 output

V2 analysis documents can be checked with the dependency-free validator:

python3 scripts/validate_analysis.py /absolute/path/to/analysis.md

The validator checks the versioned Front Matter contract, fixed required sections, evidence and confidence semantics, cognitive-candidate boundaries, claim-matrix structure, multimodal evidence, obvious Grade D contradictions, local-path and credential leaks, relative links, and Markdown whitespace. It does not judge whether the substantive analysis is correct.

Run the repository tests with:

python3 -m unittest discover -s tests -p "test_*.py"

The test suite also copies the packaged Skill into a temporary clean Codex home and runs its bundled validator there. This catches missing files and repository-only path assumptions that ordinary source-tree tests can miss.

Before creating a release tag, verify:

1. the full unit test suite passes on supported Python versions;
2. VERSION, SKILL.md, README.md, and CHANGELOG.md name the same release;
3. the packaged validator accepts tests/fixtures/valid-v2.md after a clean copy;
4. the working tree contains no duplicate or generated files;
5. the documented pinned installation command succeeds from the release tag.

Repository layout

CHANGELOG.md

skills/video-link-analysis/
├── VERSION
├── SKILL.md
├── agents/
│   └── openai.yaml
├── references/
│   ├── evidence-grading.md
│   ├── claim-evidence-protocol.md
│   ├── test-cases.md
│   └── visual-analysis-protocol.md
├── scripts/
│   └── validate_analysis.py
└── templates/
    └── analysis-template.md

scripts/
└── validate_analysis.py

tests/
├── fixtures/
│   ├── valid-v2.md
│   ├── v22-full-transcript.md
│   ├── v22-hard-subtitle.md
│   ├── v22-multiple-cognition.md
│   ├── v22-no-cognition.md
│   └── v22-unverified.md
├── test_clean_install.py
├── test_skill_structure.py
└── test_validate_analysis.py

中文说明

video-link-analysis V2.2 的定位是 Source Analyzer（来源分析器）：把视频链接转化为可核查、可追溯、可进入个人知识系统的多模态 Markdown 来源分析稿，而不是只生成内容摘要，也不替用户生成最终认知。

它会根据视频类型自适应核对字幕、章节、口播、关键画面、图表和屏幕演示。文档稳定区分来源事实、作者观点、Skill 推断与外部核验结论；认知候选明确标记为 AI 建议且尚未被用户确认。输出通过 video-analysis/v1 Front Matter 与 Personal Knowledge System 兼容，但当前仍由用户手动上传，不调用其 API 或数据库。

使用示例：

使用 $video-link-analysis 分析这个视频，自适应核对口播、字幕和关键画面，
报告证据等级、分析置信度与覆盖范围，把关键命题追溯到视频和外部证据，
进行事实核验和反方论证，并保存成 Markdown：
https://example.com/video/123

V1 生成的 Markdown 文档仍可继续使用；V2 验证器只针对采用新 Frontmatter 和多模态证据结构的文档。

License

MIT