{
  "video": {
    "id": "LC3-P7v3yoI",
    "title": "Skills are the New SDKs - Elvin Aghammadzada, DataRobot",
    "duration": 1600,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "Online"
  },
  "analysis": {
    "video_id": "LC3-P7v3yoI",
    "title": "Skills are the New SDKs - Elvin Aghammadzada, DataRobot",
    "one_liner": "A DataRobot engineer argues that because context is a budget — performance degrades past ~25-40% of the window — skills, not more MCP tools, are the right unit of packaging: progressive disclosure makes them ~10x cheaper than tool definitions, and \"teachability\" becomes the new enterprise moat.",
    "summary": "Aghammadzada attacks the \"beautiful lie\" of infinite context windows, citing the context rot paper showing degradation after 25% of the window is used and a practical \"smart zone\" that ends around 40% — which you can blow through with MCP tool definitions before the user has typed anything. Skills fix this through progressive disclosure across three levels (sub-100-token front matter acting like a database index, a sub-5K-token markdown body loaded on activation, and level-three scripts whose output alone returns to context). Strategically he reframes moats: the SaaS-era moat was friction (high switching cost), but since agents can now port a million lines from Python to Rust, the new moat is fluency — how fast a new agent harness picks up your operational knowledge, which he calls teachability. A co-presenter (Carson) adds that skills and MCP interlock rather than compete, and the talk closes on the risks: prompt injection, no process isolation, unverified marketplaces, and research showing LLM-written skills actually hurt performance.",
    "key_points": [
      "Every AI app has three layers — the UI the user sees, the system prompt and tool descriptions the model sees, and the schema/tool-call I/O the data sees — and the bugs live in the second and third, hidden from the user.",
      "The context rot paper found performance starts degrading after ~25% of the context window is consumed (256K of a 1M window); in practice he draws the line at 40%, past which you enter the \"dumb zone\".",
      "Fifteen MCP servers can burn over 100,000 tokens per session in tool definitions alone, before the conversation starts; skills cover the same operational ground at least 10x smaller thanks to progressive disclosure.",
      "Traffic to documentation websites from coding agents jumped from 10% to 50% in a year — docs are written for humans who can ask a follow-up or Google it, a capability models lack without a context engine around them.",
      "Moat thesis: the old SaaS moat was friction (make leaving expensive), but Claude Code rewriting a million lines from Python to Rust has collapsed switching costs; skills create a fluency moat, which is offensive rather than defensive and compounds with each skill added.",
      "\"Teachability\" is proposed as a new item on the enterprise evaluation checklist alongside security, compliance, data governance, SLAs, logging, tracing and auth integration — how fast a new agent harness can pick up your encoded operational knowledge and get a result in seconds.",
      "A skill has three levels: front matter (typically under 100 tokens, always in the system prompt, acting like a DB index), the markdown body (under 5K tokens, loaded only on activation), and scripts — either executable (Claude Code runs get_deployment_features.py and returns only the output) or illustrative example code.",
      "Skills vs MCP: both are Anthropic's and not competing. MCP defines what actions to take and gives isolation, auth, and hosted horsepower (GPUs, semantic search over hundreds of terabytes); skills change how the agent thinks, can self-modify, and run on the agent's own machine. A skill can even expose, dump, or spin up an MCP server.",
      "Ecosystem: 26+ platforms support skills (Claude Code, Codex, Copilot, Gemini CLI), there are hundreds of thousands of skills and paid marketplaces, and the flagship open-source agent-skills repo — with sub-agents for context hygiene and self-evolving skills — is a coding agent with roughly 10 tools that can do anything given the right skills.",
      "Recently published research found LLM-generated skills hurt performance: they use more tokens and spend more time reasoning about the problem rather than making the agent faster or leaner on context.",
      "Risk list: prompt injection from publicly published skills, no process isolation (everything runs on your laptop, unlike MCP), and marketplaces with no verification control — \"just like how NPM was scary 10 years ago\"."
    ],
    "takeaways": [
      "Budget context like a limited resource: keep total context under ~40% before the first user turn, which means not attaching 15 MCP servers' worth of tool definitions by default.",
      "Replace tool sprawl with a skills/capabilities folder on top of one good general-purpose agent — build skills for the domain, not a new agent per domain, as long as the base reasoning model is strong.",
      "Write skills so the front matter is a tight sub-100-token index and the body stays under 5K tokens; push heavy work into level-three scripts so only the execution output enters context.",
      "Have humans write the skills, not the LLM — the research shows generated skills cost more tokens and more reasoning time, and \"a skill is only as good as the human who wrote it\".",
      "Treat skills as software: version them, evaluate them, and test them, and vet marketplace skills by author and stars the way you once vetted NPM packages.",
      "Keep MCP for what it's actually good at — authentication, restricted-environment data access, process isolation, and remote horsepower — and use a skill to do progressive disclosure for that MCP server."
    ],
    "topics": [
      "skills",
      "context-engineering",
      "mcp",
      "agents",
      "progressive-disclosure",
      "enterprise-ai",
      "developer-tools",
      "ai-security"
    ],
    "tools": [
      "Claude Code",
      "Anthropic",
      "MCP (Model Context Protocol)",
      "Agent Skills",
      "Codex",
      "GitHub Copilot",
      "Gemini CLI",
      "DataRobot",
      "Claude Agent skills registry",
      "React",
      "NPM",
      "Python",
      "Rust"
    ],
    "quotes": [
      {
        "text": "The first one is that beautiful lie that we've been told that the latest frontier models have infinite context windows.",
        "at": "00:57",
        "url": "https://www.youtube.com/watch?v=LC3-P7v3yoI&t=57s"
      },
      {
        "text": "At one point, makes you feel bad that Cloud Code is making the same mistake it made 5 minutes ago.",
        "at": "02:56",
        "url": "https://www.youtube.com/watch?v=LC3-P7v3yoI&t=176s"
      },
      {
        "text": "If it's connected to 15 MCP server, I'm pretty sure it's consuming over 100,000 tokens per session just in tool definitions itself. Without starting the conversation yet.",
        "at": "13:46",
        "url": "https://www.youtube.com/watch?v=LC3-P7v3yoI&t=826s"
      },
      {
        "text": "A skill is only as good as the human who wrote it based on our experience.",
        "at": "26:19",
        "url": "https://www.youtube.com/watch?v=LC3-P7v3yoI&t=1579s"
      }
    ],
    "words": 4924
  },
  "summary_url": "/#LC3-P7v3yoI",
  "transcript": {
    "html": "/transcripts/LC3-P7v3yoI.html",
    "txt": "/transcripts/LC3-P7v3yoI.txt",
    "vtt": "/transcripts/LC3-P7v3yoI.vtt"
  }
}