{
  "video": {
    "id": "Yn8h5Ip-L9c",
    "title": "Do you even need human review?",
    "duration": 119,
    "upload_date": "2026-07-20",
    "channel": "Matt Pocock",
    "source": "https://www.youtube.com/watch?v=Yn8h5Ip-L9c",
    "event_format": ""
  },
  "analysis": {
    "video_id": "Yn8h5Ip-L9c",
    "title": "Do you even need human review?",
    "one_liner": "Matt Pocock's framing for AI-assisted engineering: don't hand every checkpoint to AI at once — shift the human review checkpoint left (align at planning) for big work, and right (review just before it goes live) for small diffs.",
    "summary": "Pocock argues the best way to automate engineering work is to figure out precisely where you still need human review, rather than converting the whole pre-AI plan → ship → review chain to AI in one go. For very large pieces of work you should be involved at the initial planning stage and align with the AI first (what his Grill Me, Grill with Docs and Wayfinder skills are for), plus keep a human review before production. For small, low-impact changes you can skip the alignment entirely, specify it in a single prompt, and only review at the end — and for things like a pure refactor or an internal documentation rename you can align early and let the AI ship straight to prod. The organising language is shifting the checkpoint left for big work and right for small work.",
    "key_points": [
      "The pre-AI process is plan work → ship work → review work, and every checkpoint in it is human; the common dream is translating all of them to AI at once, which Pocock says doesn't make sense.",
      "On very large pieces of work you want to align with the AI first, at the initial planning stage — that's what his Grill Me, Grill with Docs and Wayfinder skills do — and keep human review right at the end before shipping to production.",
      "Some work needs little or no review at the end: a refactor, or an internal documentation rename — the human aligns early, considers the plan done, and the AI ships all the way to prod.",
      "Other work needs no alignment at all: specify it in a single prompt, don't bother grilling, let it ship through the process, and review only before it goes live.",
      "The no-alignment path suits very simple changes and very small diffs that won't make a big splash — changing the color of a button, changing the name of a title, or a small bug fix that's a couple of lines.",
      "The stated mechanism: the smaller the diff, the easier it is to totally regenerate the code and hold it all in context; for massive work you align beforehand because fixing it at a later stage is really expensive.",
      "The vocabulary he recommends: shift checkpoints left for big work, shift them right for small work."
    ],
    "takeaways": [
      "Stop treating 'automate it all' as the goal — audit your plan/ship/review checkpoints individually and decide which ones still need a human.",
      "Size the work first, then place the checkpoint: big work gets a human at planning (align/grill before any code), small work gets a human only at the final pre-production review.",
      "For a couple-of-lines bug fix, a button colour or a title rename, skip alignment entirely and just write one prompt — the alignment overhead isn't worth it.",
      "For low-risk mechanical work like refactors and internal doc renames, consider letting the AI ship to prod after early alignment, with no end review.",
      "Front-load scrutiny on large work specifically because late correction is expensive, and small diffs can just be regenerated wholesale in context."
    ],
    "topics": [
      "agents",
      "human-in-the-loop",
      "code-review",
      "ai-coding-workflow",
      "planning",
      "context-management",
      "skills"
    ],
    "tools": [
      "Grill Me",
      "Grill with Docs",
      "Wayfinder"
    ],
    "quotes": [
      {
        "text": "Everyone's obsessed with automating work with engineers. And the way you do it best is to figure out where you need human review.",
        "at": "00:00",
        "url": "https://www.youtube.com/watch?v=Yn8h5Ip-L9c"
      },
      {
        "text": "A lot of people's dream, I think, is to have these all translated over to AI at once. But for me, that doesn't really make sense.",
        "at": "00:21",
        "url": "https://www.youtube.com/watch?v=Yn8h5Ip-L9c&t=21s"
      },
      {
        "text": "You shift checkpoints left for big work, for stuff you need to align on first, and you can shift it to the right for small work.",
        "at": "01:33",
        "url": "https://www.youtube.com/watch?v=Yn8h5Ip-L9c&t=93s"
      },
      {
        "text": "The reason for that is that the smaller the diff, the easier it is to totally regenerate the code and hold it all in context.",
        "at": "01:40",
        "url": "https://www.youtube.com/watch?v=Yn8h5Ip-L9c&t=100s"
      }
    ],
    "words": 500
  },
  "summary_url": "/#Yn8h5Ip-L9c",
  "transcript": {
    "html": "/transcripts/Yn8h5Ip-L9c.html",
    "txt": "/transcripts/Yn8h5Ip-L9c.txt",
    "vtt": "/transcripts/Yn8h5Ip-L9c.vtt"
  }
}