{
  "video": {
    "id": "_ehJyfHg1Vk",
    "title": "The Era of Compound Engineering — Kieran Klaassen, Every/Cora",
    "duration": 1238,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer"
  },
  "analysis": {
    "video_id": "_ehJyfHg1Vk",
    "title": "The Era of Compound Engineering — Kieran Klaassen, Every/Cora",
    "one_liner": "Kieran Klaassen built and runs the Cora AI email client as a solo engineer without writing a line of code this year, by spending half his time teaching a repo-resident memory system his judgment and taste so the planning-building-reviewing middle can run unattended for hours.",
    "summary": "Klaassen traces two years of moving bottlenecks — bad code, then plans, then deciding what to build — to the conclusion that implementation is basically solved and judgment and taste are what remain scarce. His answer is 'compound engineering': a loop (brainstorm → plan → work → review → polish → compound) where the human is 'the bread' at both ends and the AI runs the middle, with every correction extracted into solution documents stored in the repository so the same lesson never has to be given twice. He demos his open-source compound engineering plugin's commands — /ce ideate, /ce doc review, /ce brainstorm, /LFG, /ce polish, /ce compound — and argues a single engineer with a compounding system beats whole teams using AI without one.",
    "key_points": [
      "He has not written a single line of code this year, has barely looked at most of it, and still solo-ships Cora — a complete, agent-native AI email inbox used by thousands of people, Rails backend and React frontend, with the v2 rebuild started in January; he has design and hard-database support but no engineering team.",
      "He is an ex-VP of engineering and founder who deliberately did the opposite of hiring when Sonnet 3.5 landed, to see how far AI could go before he needed to grow a team.",
      "The bottleneck kept moving: first bad code and hallucination (fixed with agents, skills, review), then the plan, then deciding what to build; when he found himself repeating the same corrections he moved memory out of a CLAUDE.md that had grown too large into a system that remembers.",
      "The loop is brainstorm, plan, work, review, polish, compound, repeat — the 'human AI sandwich' where the brain is on at both ends (choosing the problem, and raising the bar at the end) and the AI runs the middle unattended, overnight and in parallel.",
      "His explicit rule: 50% of your time on shipping the feature, 50% on teaching the system whatever it got wrong — and against the 'but tokens' objection, he says storing the answers as solution documents in the repo is more token-efficient long term because you skip review, correction and deep research.",
      "The compound engineering plugin is open source, installs in Codex, Claude Code, Cursor 'plus 10 others', and he says hundreds of thousands of people now use it daily; it was never designed as hype, just his own working plugin.",
      "Demos: /ce ideate reads open tickets across Linear, GitHub, Slack and Intercom, argues about what's worth doing, scores ideas against a strategy doc or OKRs (made with /ce strategy) and outputs a shareable HTML page; /ce doc review returns sharp questions on a PRD; /ce brainstorm is a blocked-off, brain-on session that asks deliberately few questions — he criticizes other libraries for over-questioning ('very easy to get 30 questions and feel, wow, I did so much'); /LFG runs for hours doing planning, work, review and testing, opens a PR, dogfoods, and attaches before/after video and screenshots.",
      "/ce polish is not QA but bar-raising — he runs it in Cursor with the change explained on the left and the product on the right; his example was a page rendering the logo mark twice, which he fixed and then fed to /ce compound so future design work is tagged with that rule.",
      "He argues you should document the thinking, not the code — deliberately 'anti-developery' — because reasoning traces let a postmortem ask which decision, by whom or which agent, led to an incident, and turn that into a behavior change."
    ],
    "takeaways": [
      "Spend 50% of every interaction making the system better next time, not just shipping the feature: whenever you repeat yourself, extract it into stored knowledge so it never has to be said again.",
      "Invest in the middle of the loop until it is boring and runs without you — do it manually, feel where it's off, iterate; you know you're there when you can start a three-hour run and the result is always good.",
      "Keep your brain on at the ends: don't offload the problem definition to the AI, and at the end raise the bar rather than fixing bugs — if things are broken at the polish step, the automated run failed.",
      "Store the reasoning and solution documents inside the repository (not just the code), so plans, brainstorms and postmortem learnings are in-context for the next run.",
      "Adopt the standard that the next feature should be easier because you shipped this one — the inversion of normal complexity accretion.",
      "You don't need his plugin: build your own compounding knowledge base, even if it's just storing information in files."
    ],
    "topics": [
      "compound engineering",
      "agents",
      "ai coding",
      "agent memory",
      "solo founder",
      "workflow",
      "code review",
      "taste and judgment"
    ],
    "tools": [
      "Cora",
      "Every",
      "compound engineering plugin",
      "Claude Code",
      "Codex",
      "Cursor",
      "Claude Sonnet 3.5",
      "Claude Design",
      "CLAUDE.md",
      "MCP",
      "Rails",
      "Ruby",
      "React",
      "Linear",
      "GitHub",
      "Slack",
      "Intercom",
      "PowerPoint"
    ],
    "quotes": [
      {
        "text": "I haven't written a single line of code this year.",
        "at": "00:22",
        "url": "https://www.youtube.com/watch?v=_ehJyfHg1Vk&t=22s"
      },
      {
        "text": "I see that one engineer with a compounding system just beats teams like full teams that use AI that don't.",
        "at": "06:27",
        "url": "https://www.youtube.com/watch?v=_ehJyfHg1Vk&t=387s"
      },
      {
        "text": "It's kind of the human AI sandwich where the human is the bread and the AI is the middle part.",
        "at": "06:54",
        "url": "https://www.youtube.com/watch?v=_ehJyfHg1Vk&t=414s"
      },
      {
        "text": "The bet is implementation is only getting cheaper and judgment is not.",
        "at": "18:50",
        "url": "https://www.youtube.com/watch?v=_ehJyfHg1Vk&t=1130s"
      },
      {
        "text": "The next feature should be easier to build because you ship this one.",
        "at": "19:47",
        "url": "https://www.youtube.com/watch?v=_ehJyfHg1Vk&t=1187s"
      }
    ],
    "words": 3697
  },
  "summary_url": "/#_ehJyfHg1Vk",
  "transcript": {
    "html": "/transcripts/_ehJyfHg1Vk.html",
    "txt": "/transcripts/_ehJyfHg1Vk.txt",
    "vtt": "/transcripts/_ehJyfHg1Vk.vtt"
  }
}