shanezhong/youtube-ai-interview-finder

YouTube AI Interview/Podcast Finder

Automates the source-discovery step of the substack-writer content pipeline. Hand-curates one YouTube video per run, matching editorial criteria for a blog post.

When to use

• Daily cron at the start of a content pipeline that ends in Substack + Medium publishing
• Ad-hoc: "find me a video for today's blog"
• Anywhere you'd otherwise have to manually scan YouTube for a usable interview/podcast

Editorial Criteria (hard-coded)

A candidate passes ONLY if all are true:

1. Human-to-human conversation — interview or podcast format. NOT a promotional/explainer/review video. NOT "AI doing the interview" (e.g. "I let AI interview me for a job").
2. Duration — between 10 minutes and 3 hours.
3. Upload date — current month (e.g. April 2026 if running in April 2026).
4. Topic relevance — about AI as substantive content. Reject:
   • Generic "what is AI / what is LLM" explainers
   • AI safety doom without actionable content
   • AI hiring/interview-tool reviews
   • Job-prep mock interviews
5. Editorial value — prefer videos with actionable content: someone explaining HOW they use AI for X, concrete playbooks, real workflows, real tradeoffs. Practitioner-led > pundit-led.
6. Not previously used — must not appear in used_videos.json.

Dedup discipline (important)

A video is added to used_videos.json ONLY after the full downstream pipeline (blog written + Substack published) completes successfully. Picking a video as the day's candidate does NOT mark it used. This means a runner-up from yesterday is still eligible today.

Workflow

Step 1 — Run the search

When this skill loads, the system tells you the base directory (e.g. Base directory for this skill: /Users/shane/.claude/skills/youtube-ai-interview-finder). Use that to construct the script path. The env vars $CLAUDE_PLUGIN_ROOT and $CLAUDE_SKILL_DIR are not reliably set for locally-installed skills — always derive from the base directory.

/Users/shane/Documents/playground/.venv/bin/python \
  <base_directory>/scripts/find_video.py

For example, when locally installed:

/Users/shane/Documents/playground/.venv/bin/python \
  /Users/shane/.claude/skills/youtube-ai-interview-finder/scripts/find_video.py

This:

1. Reads APIFY_TOKEN from /Users/shane/Documents/playground/.env
2. Runs Apify actor streamers/youtube-scraper with 3 search queries: "AI interview", "AI podcast", "interview AI founder"
3. Filters by duration (10min–3hr), drops any video already in used_videos.json
4. Writes:
   • ai_writing/auto_youtube_finder/candidates_raw_{YYYY-MM-DD}.json — full Apify dump
   • ai_writing/auto_youtube_finder/candidates_{YYYY-MM-DD}.json — deduped + duration-filtered shortlist

Apify's uploadDate=month filter is loose — it returns videos older than a month too. The agent must filter strictly to the current month in Step 2.

Step 2 — Agent picks one

Read candidates_{YYYY-MM-DD}.json. Apply the editorial criteria above. For each candidate, evaluate the title + description + channel to judge:

• Is this actionable practitioner content, or doom/explainer pundit content?
• Is the upload date in the current month? (Apify's filter is unreliable — re-check date field strictly.)
• Does the content match Shane's substack audience (solo dev / AI tools / data / business builder)?

Pick one winner. Pick one runner-up (in case the winner fails downstream). Write a JSON file with explicit reasoning:

{
  "date": "YYYY-MM-DD",
  "pick": {
    "id": "...",
    "title": "...",
    "url": "...",
    "channel": "...",
    "guest": "...",
    "duration": "HH:MM:SS",
    "viewCount": 0,
    "date": "YYYY-MM-DD"
  },
  "reasoning": "Why this video — must be specific, citing the channel's reputation, the guest's track record, and the actionable value. No filler.",
  "criteria_check": {
    "human_to_human": true,
    "is_interview_or_podcast": true,
    "duration_over_10min": true,
    "this_month": true,
    "actionable_advice": true,
    "not_generic_ai_explainer": true,
    "not_used_before": true
  },
  "runner_up": { "id": "...", "title": "...", "why_runner_up": "..." },
  "rejected_with_reasons": { "<video_id>": "<one-line reason>", ... },
  "stats": { "raw_apify_results": 0, "this_month_only": 0, "passing_all_criteria": 0 }
}

Save as ai_writing/auto_youtube_finder/pick_{YYYY-MM-DD}.json.

Step 3 — Hand off (optional)

If invoked as part of the daily content cron, emit the picked video URL to stdout / Discord, then trigger substack-writer with that URL as input.

After substack-writer finishes successfully (Substack post published), append the video ID to used_videos.json:

{ "video_ids": ["wc8FBhQtdsA", "..."], "_doc": "..." }

If substack-writer fails, do NOT append — the video is still eligible tomorrow.

File locations

| Path | Purpose | |------|---------| | <skill_base_dir>/scripts/find_video.py | Apify caller (skill_base_dir given on skill load) | | ~/Documents/playground/.env (APIFY_TOKEN) | Auth | | ~/Documents/playground/ai_writing/auto_youtube_finder/used_videos.json | Dedup ledger (append after publish) | | ~/Documents/playground/ai_writing/auto_youtube_finder/candidates_raw_{date}.json | Raw Apify dump | | ~/Documents/playground/ai_writing/auto_youtube_finder/candidates_{date}.json | Deduped shortlist | | ~/Documents/playground/ai_writing/auto_youtube_finder/pick_{date}.json | Picked video + reasoning |

Dependencies

• Python 3.11 venv at /Users/shane/Documents/playground/.venv
• apify-client, python-dotenv
• APIFY_TOKEN in .env

Install if missing:

/Users/shane/Documents/playground/.venv/bin/pip install apify-client python-dotenv

Common rejection patterns (for agent reference)

| Title pattern | Reason to reject | |---------------|------------------| | "I Let AI Interview Me for X" | AI is the interviewer, not the topic | | "AI Job Interview Questions & Answers" | Job-prep niche, not editorial content | | "Godfather of AI Warns / Doom / We've Lost" | Doom without actionable value | | "Why AI CEOs Are Building Bunkers" | Pundit doom, not practitioner | | "What is AI / Beginner's Guide to LLMs" | Generic explainer | | Channel: news outlet (CBS / BBC / Bloomberg short clips < 20 min) | Not enough depth for blog |

Common acceptance patterns

| Pattern | Why | |---------|-----| | Lenny's Podcast / Dwarkesh Patel / a16z / Latent Space | Practitioner-focused, deep | | Guest is a builder/founder talking about HOW they build | Actionable | | 60-120 min long-form interview | Enough material for a blog | | Title hints at concrete claim (numbers, playbook, specific company) | Substantive |