Stop summarizing videos into a wall of bullet points. Start producing structured knowledge.
Why this exists. Generic AI summarizers flatten every video into the same shape: a paragraph and some bullets. That's fine for a single watch — but it loses the structure of a tutorial, can't cross-reference the GitHub repo a tutorial points to, and forgets what packages were mentioned the moment the conversation ends.
youtube-analyzerproduces format-aware, structured markdown you can re-read, search, and feed into other tools.
flowchart LR
A[YouTube URL] --> B[Phase 1<br/>Transcript fetch<br/>4-tier fallback]
B --> C[Phase 2<br/>Detect content type<br/>+ interactive config]
C --> D[Phase 3<br/>Token-aware partitioning<br/>+ optional GitHub clone]
D --> E[Phase 4<br/>Parallel format agents<br/>+ synthesis]
E --> F1[Structured markdown<br/>document]
E --> F2[Inline chat<br/>analysis]
style A fill:#1a1f3a,stroke:#5eead4,color:#fff
style F1 fill:#1a1f3a,stroke:#a78bfa,color:#fff
style F2 fill:#1a1f3a,stroke:#a78bfa,color:#fff
| Format-aware | Detects whether the video is a tutorial, course, finance video, interview, lecture, or general — and dispatches the matching analysis workflow. |
| Multi-agent on long videos | Token-aware partitioning splits 100K-token videos across parallel agents, then synthesizes the chunks into one document. |
| GitHub repo cross-reference | For tutorials, optionally clones the linked repo and produces Mermaid diagrams of structure, dependencies, and patterns — so you can see what the video taught vs. what the actual code does. |
| Repository handoff | Preserves verified GitHub repository links from the video, description, transcript, and supporting links with their sources, so a later explicit request can star the right repositories through the separate stars skill. |
| Package version drift | Tracks every package mentioned in tutorial videos in a local database, then queries npm / PyPI to flag versions that have moved on since the video shipped. |
| Two delivery modes | --document writes a permanent markdown file to your configured output dir; --chat returns the analysis inline so you can talk through it without saving. |
Inside Claude Code:
/plugin marketplace add AojdevStudio/agentic-utilities
/plugin install youtube-analyzer@agentic-utilities
Then ask Claude:
Analyze this video: https://youtube.com/watch?v=...
Or invoke explicitly with a flag:
--chat https://youtube.com/watch?v=... # inline, no file written
--document https://youtube.com/watch?v=... # save to configured output dir (default)
The skill verifies these on first run and exits cleanly if any are missing:
| Tool | Install |
|---|---|
yt-dlp |
pip install yt-dlp or brew install yt-dlp |
youtube_transcript_api |
pip install youtube-transcript-api |
bun |
curl -fsSL https://bun.sh/install | bash |
bun runs the four TypeScript helpers in scripts/ (transcript cleanup, content-type detection, token partitioning, package database).
Four phases, each with a blocking gate. The orchestrator never analyzes content directly — it routes work to specialized sub-agents.
Pulls the transcript with a 4-tier fallback: youtube_transcript_api (high quality) → yt-dlp auto-subs (medium) → metadata-only (low) → graceful exit. Cleans VTT artifacts, counts words, captures video metadata.
Auto-detects category (technology / finance / business / etc.) and format (tutorial / course / finance / interview / lecture / general) using keyword scoring on the title, description, tags, and channel. Anything below 60% confidence triggers a confirmation prompt. Then asks you what outputs you want, what depth, and any format-specific focus.
Estimates token count and decides single-agent (≤30K) vs. multi-agent partitioning (30K–100K → 2-3 agents, >100K → 4+ agents at ~25K tokens each). For tutorials with a GitHub repo URL, three Explore sub-agents (StructureExplorer, DependencyExplorer, PatternExplorer) run in parallel against the cloned repo and produce Mermaid diagrams.
Each chunk runs through the format's workflow (tutorial-workflow.md, finance-workflow.md, or general-workflow.md). One synthesis agent merges chunk outputs, applies the YAML frontmatter and section templates, and either writes a markdown file or returns the rendered output inline. For tutorials, every package mentioned lands in the package database with version-drift tracking.
On first document-mode run, the skill asks where to save analyses and persists your answer to .claude/youtube-analyzer.local.md in the current project. You can edit that file anytime:
---
output_directory: /absolute/path/to/your/notes
---The package database is at ~/.config/youtube-analyzer/package-db.json (auto-created).
youtube-analyzer/
├── .claude-plugin/
│ └── plugin.json
├── skills/
│ └── youtube-analyzer/
│ ├── SKILL.md Orchestration spine — 4 phases, blocking gates
│ ├── content-types.md Category/format keyword scoring
│ ├── package-database-schema.md
│ ├── references/
│ │ ├── source-selection.md Phase 1 mechanics (transcript fetch chain)
│ │ ├── scaling-and-repo-explore.md
│ │ ├── output-paths.md
│ │ └── output-templates.md
│ └── workflows/
│ ├── tutorial-workflow.md
│ ├── finance-workflow.md
│ ├── general-workflow.md
│ └── repo-exploration.md
├── scripts/
│ ├── clean-transcript.ts Strip VTT artifacts
│ ├── detect-content-type.ts Keyword-scored content-type detection
│ ├── partition-transcript.ts Token-aware partitioning for parallel agents
│ └── package-db.ts Track tutorial packages with version drift
├── assets/
│ └── hero.png
├── README.md This file
└── CHANGES.md Differences from the personal (PAI) version
- Local transcript-library browse mode (was in the personal version, dropped for portability)
- Sermon workflow as an optional companion plugin
- Channel-aware content-type tuning (auto-classify by channel signature)
MIT.
Part of the agentic-utilities marketplace. See CHANGES.md for differences from the original personal (PAI) version of this skill.