michailbul/laniameda-youtube-digest

laniameda-youtube-digest

Purpose: Watch a YouTube video so Michael doesn't have to. Extract what's actionable — tools, prompts, workflows, "if X then Y" mappings — and save to KB. For AI creatorship tutorials: extract model-specific techniques, prompt structures, and optionally encode new knowledge into studio skills.

Engine: Supadata API (supadata skill). Full API reference: references/supadata.md. Key decision: use Transcript for narrated content; use Extract (vision) when prompts/settings are shown on screen but not spoken aloud (e.g. ComfyUI node graphs, Midjourney UI, on-screen prompt text).

Cross-reference: This is the skill that laniameda-x-post-digest calls when it encounters YouTube links in X posts.

---

Triggers

• YouTube URL pasted
• "digest this video"
• "watch this for me"
• "extract from this video"
• "what's useful in this"
• "learn from this"
• "what does this teach"
• "does this have good prompting technique"
• "upgrade your skills from this"
• "digest this AI tutorial"

---

Extraction Priority (what matters to Michael)

• Agent orchestration — specific workflow architectures, "if X then Y" tool mappings, MCP names, Claude Code patterns
• Prompt engineering — templates with [BLANKS], copy-pasteable structures, named frameworks
• AI design systems — design tokens, UI generation prompts, exact tool names + use cases
• AI creatorship — exact prompts with parameters, generation workflows, modern techniques
• Model-specific craft — what works/doesn't in Midjourney, FLUX, Nano Banana, Kling, Runway, Seedance, etc.
• Tool workflows — "If I need X, use Y" with steps
• Visual direction craft — lighting language, camera language, composition rules that translate to prompts
• Negative findings — techniques that are hyped but don't work (equally valuable)

Golden rule: If Michael can't copy-paste and use it immediately, it's not specific enough. Generics = skip. Inspiration without technique = skip.

---

Workflow

Step 1 — Metadata

curl -s "https://api.supadata.ai/v1/youtube/video?id=<VIDEO_ID>" \
  -H "x-api-key: $SUPADATA_API_KEY"

Check description for linked prompts, swipe files, preset downloads, "link in bio" resources → fetch with web_fetch if relevant.

Step 2 — Transcript or Extract

See references/supadata.md for full API reference and curl commands.

Choose based on content type:

• Narrated content → /youtube/transcript?url=<URL>&text=true (native captions, 1 credit)
• Visual-heavy tutorial (prompts on screen, UI demos, ComfyUI graphs, settings panels) → /extract with vision schema. Use when important technique content is shown but not spoken.
• If transcript returns no content → retry without lang param, or fall back to Extract.

# Transcript (default)
curl -s "https://api.supadata.ai/v1/youtube/transcript?url=<URL>&text=true" \
  -H "x-api-key: $SUPADATA_API_KEY"

# Extract (for visual content — async, returns jobId)
curl -X POST "https://api.supadata.ai/v1/extract" \
  -H "x-api-key: $SUPADATA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "url": "<URL>",
    "prompt": "Extract all AI image prompts shown on screen. Include exact text, tool/platform visible, and parameter settings.",
    "schema": { ... }
  }'

Step 3 — Fast verdict (always deliver this first)

[VIDEO TITLE] — [Channel] (~Xmin)

What it is: [one sentence — what they show + tools used]

Verdict: USEFUL / PARTIAL / SKIP

Why:
• [specific reason]
• [what's new]
• [what's missing]

Interest match: [which of Michael's interests]
Repurposable: Yes / No

Preview:
• [Tool/Technique]: [what it does / one line]
• [Prompt/workflow]: [brief]

Skill impact: [which existing skill this would update, or "new skill warranted", or "none"]

Digest and save? (yes / skip / partial)

Step 4 — Deep digest (on approval)

Extract ONLY:

• Specific tool names
• Exact prompts shown or described (copy-paste ready)
• Step-by-step workflows
• "If X then Y" mappings
• Code snippets / commands
• Model-specific rules ("in Kling v2.5, X causes Y")
• Tool settings, parameters, version numbers
• Technique verdicts (works / doesn't work / when to use)
• Direct quotes if they capture a principle cleanly

Skip: channel intros, sponsor segments, generic motivation, "AI is amazing" takes, vague inspiration.

Step 5 — Save to KB

Location: ~/work/laniameda/laniameda-hq/content-kb/sources/youtube/YYYY-MM-DD-<slug>/

Files:

• meta.json — title, url, channel, date, verdict, tools_mentioned, models_mentioned, techniques, tags
• digest.md — full extraction
• prompts.md — prompts/templates only
• workflows.md — step-by-step workflows only
• techniques.md — model-specific rules, technique verdicts, what works vs what doesn't

If marketing-ready → also save to content-kb/marketing-ready/YYYY-MM-DD-<slug>.md

Step 6 — Skill encoding check

Read references/skill-upgrade-protocol.md for the full decision tree.

Short version: After saving to KB, ask: does this video teach something that isn't already in our skills?

• Yes, it refines an existing skill → propose a patch to the relevant SKILL.md, ask Michael to approve before writing
• Yes, it warrants a new skill → propose skill name + one-line purpose, ask Michael to approve, then build it using skill-creator
• No, KB save is enough → done

Always ask before writing to any SKILL.md or creating a new skill folder.

---

Hard Rules

• Deliver fast verdict first, always — don't dump full transcript unprompted
• If video is mostly generic motivation/inspiration → SKIP, say why in one line
• SKIP verdict if video is: generic AI hype, tool demos with no technique depth, "10 prompts you need" lists with no explanation of why they work
• Never summarize in vague terms — specifics or nothing
• Save the actual content, not a paraphrase of what the video "is about"
• Never paraphrase technique — extract the actual rule, the actual prompt, the actual finding
• Model specificity matters — "works in Midjourney" ≠ "works in FLUX" — always tag the model