YouTube Research

Seed question: What do practitioners actually DO (not just document)?

Extract structured knowledge from YouTube video transcripts. Videos capture practitioner insights, tips, and patterns that don't appear in formal documentation.

Phases

Phase 1: Discovery

Method A: Web Search

WebSearch: "[topic] tutorial 2025 youtube"
WebSearch: "[topic] best practices youtube"
WebSearch: "[topic] tips advanced youtube"

Method B: Playwright Discovery (Better for Recommendations)

1. Navigate to YouTube search
2. Search for [topic] tutorial [year]
3. Capture video titles and URLs from results
4. Click into a relevant video to get recommendations
5. Capture recommended videos (YouTube's algorithm surfaces related content)

Method C: Known Creator Channels

• Official company channels (Anthropic, Google, Microsoft)
• Tech educators (Fireship, NetworkChuck, Traversy Media)
• Conference channels (AI Engineer, React Conf)

Phase 2: Transcript Acquisition

Delegates to video-transcript-extraction skill.

For each discovered video:

1. Invoke video-transcript-extraction
2. If transcripts disabled, skip and note in research log
3. Save transcript to working directory

Phase 3: Pattern Extraction

For each transcript, extract:

Structural Elements:

• Concepts defined (new terms, mental models)
• Workflows described (step-by-step processes)
• Anti-patterns mentioned (what NOT to do)
• Tips and tricks (practitioner shortcuts)
• Tools/libraries mentioned (ecosystem components)

Quality Signals:

• Confidence markers: "Always do X" vs "I prefer X"
• Source authority: official channel vs creator opinion
• Recency: check video date, concepts may be outdated
• Triangulation: same pattern from 3+ sources = high confidence

Phase 4: Synthesis

Cross-Video Analysis:

1. Group similar concepts across transcripts
2. Identify consensus patterns (3+ sources agree)
3. Flag contradictions for human review
4. Note unique insights from single sources (lower confidence)

Output Structure:

# [Topic] YouTube Research Analysis

**Date:** YYYY-MM-DD
**Videos Analyzed:** N

## Videos Analyzed
| Video | Creator | Focus | Key Value |

## New Patterns Discovered
## Reinforced Patterns
## Contradictions Found
## Methodology Notes

Phase 5: Export

Normal mode:

output/
  HANDOFF.md          — Summary + suggested next actions
  corpus.json         — Machine-readable (auto-generated)
  all_content.md      — Consolidated markdown
  analysis/
    overview.md
    patterns.md
    themes.md

Budget mode: Skip corpus.json unless user requests. HANDOFF.md and all_content.md always generated.

Relational memory: Memorize key findings if relational-memory MCP is configured (skip if not available):

mcp__relational-memory__memorize(
    agent_name="youtube-research",
    layer="recent",
    content="[key finding]",
    metadata={"topic": "...", "videos": N}
)

Detail Levels

Choose detail level BEFORE starting extraction:

Level	Mode	Per-Video Output	When to Use
0-3	Quick	1-2 sentences, topic tags	Triage many videos, initial discovery
4-6	Balanced	Summary + key points + notable quotes	Standard research, known-good sources
7-10	Deep	Full extraction, timestamps, cross-references	High-value topics, building corpus

Adjusting mid-research: Start at 4-6 by default. Increase to 7-10 if finding gold. Decrease to 0-3 if hitting diminishing returns.

Self-Managing Iteration

This skill is SELF-MANAGING. No user input needed for iteration decisions.

Budget-Aware Self-Review (After Each Pass)

## SELF-REVIEW - Pass N

### Value Assessment
1. Patterns found this pass: [count]
2. Novel insights (not seen before): [count]
3. Reinforced patterns: [count]

### Budget Check
4. Detail level used: [N]
5. Token investment: HIGH/MEDIUM/LOW
6. Value delivered: HIGH/MEDIUM/LOW
7. Value/Token ratio: GOOD/ACCEPTABLE/POOR

### Decisions
8. Continue? [YES/NO]
9. Adjust detail level? [UP/DOWN/SAME]

Decision Algorithm

def should_continue():
    consecutive_low = count_trailing_lows(pass_history)
    if consecutive_low >= 2:
        return STOP, "Data exhausted"
    if saturation == "YES":
        return STOP, "Saturation confirmed"
    if value_token_ratio == "POOR" and consecutive_low >= 1:
        return STOP, "Diminishing returns"
    return CONTINUE, "Proceed to next pass"

Saturation Detection

• NO: New concepts, diverse sources
• BEGINNING: Meta-themes repeating, still finding techniques
• YES: Multiple videos saying same things, no novel patterns

Topic-Based Escalation

Consult reference/topic-based-escalation.md when extracted content touches:

• Safety/trust claims → suggest DIP
• Power structures / "who benefits" → suggest cui-bono
• Geopolitical/military → suggest DIP + cui-bono lenses
• Contrarian single-source → suggest dialectic-spiral (full) + iterative-verification

Mid-research escalation is a suggestion, not automatic. Note it and let the user decide.

Budget Mode

Activation (any of these):

1. Explicit flag: --budget or -b
2. Auto-detect: If budget-mode skill is active in session
3. Inherited: If invoked from research hub in budget mode

When active:

• Detail levels 3-5 instead of default 4-6
• Cap dialectic at 2 rounds if escalating
• corpus.json is OPT-IN (ask user, default: no)
• Relational-memory memorize still happens (cheap, always valuable)

Note: After context compaction, auto-detection may fail. Re-invoke budget-mode skill or pass --budget explicitly.

Propagation: When invoking other skills, pass budget: "Invoking video-transcript-extraction --budget"

Tips

1. Timestamps matter: Note timestamps for verification of specific claims
2. Creator bias: Sponsored content may promote specific tools
3. Recency decay: Tech videos older than 18 months may have outdated info
4. Comment gold: Video comments sometimes have better tips than the video (not accessible via transcript)
5. Playlist structure: Tutorial series may build concepts progressively

Limitations

• Cannot access videos with disabled transcripts (fallback to Whisper via video-transcript-extraction)
• Auto-generated transcripts may have errors
• Cannot see visual demonstrations (code on screen, diagrams)
• Comments and community notes not accessible
• Live streams may have transcript issues

Cross-References

• video-transcript-extraction — called for Phase 2 (transcript acquisition)
• deep-investigation-protocol — escalation target for trust/safety topics
• cui-bono — escalation target for power analysis topics
• cui-bono + financial-mcp — escalation target when financial dimension needed (a dedicated stonk agent is in design — issue #61)
• dialectic-spiral — used when findings need stress-testing
• reference/topic-based-escalation.md — shared escalation logic

Vasana

A vasana is a pattern that persists across unrelated contexts. If during this task you notice such a pattern emerging, it may be worth capturing. This skill works best alongside the vasana skill and vasana hook from the Vasana System plugin.

Modify freely. Keep this section intact.