{
  "video": {
    "id": "7jjudsEhBtM",
    "title": "Skills are new features: Building Skill-Centric Harness — Yogendra Miraje, FactSet",
    "duration": 1044,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "7jjudsEhBtM",
    "title": "Skills are new features: Building Skill-Centric Harness — Yogendra Miraje, FactSet",
    "one_liner": "A FactSet principal AI engineer shows how to add Anthropic-style skills to your own agent harness — registry, system prompt, file-read tool — and why, at enterprise scale, skills become the features your product ships and need routing, evals and governance like code.",
    "summary": "Yogi Miraje argues that when an agent becomes the product's interface, features stop being screens and buttons and become skills: prompts define who the agent is, tools define what it can connect to, and skills define how a task gets done — the right home for business logic. He walks through the bare-minimum harness support for skills (a skill registry, a system prompt built from name/description/path, and a file read tool, plus bash or a code sandbox for scripts) with three worked example skills, and shows how progressive disclosure and description wording drive routing. He then covers the failure modes he hit at FactSet: skills cut by data model instead of user intent, a model upgrade silently breaking skills whose critical instruction sat at the end, and system-prompt stuffing collapsing past ~10 skills. He closes on five aspects of skill-library governance — admission, ownership, boundaries, lifecycle, coherence — implemented by borrowing decades-old code practices.",
    "key_points": [
      "His previous talk at the same conference proposed 'blueprints' — a recipe of steps handed to agents so they don't rediscover their path each time; he now says that was 'simply a skill in a very naive form', and once Anthropic open-sourced skills in October there was no point maintaining their own standard, so FactSet dropped blueprints and adopted skills fully.",
      "The who/what/how framing: prompts define who the agent is, tools define what it can connect to, skills tell how a task gets done — and that is where business logic belongs. In finance, equity research and wealth management 'used to be the buttons, dropdowns and screens and now they are merely skills'.",
      "Bare-minimum skill support in your own harness is three things: a skill registry (name, description, path), a system prompt built from those entries, and a basic file read tool — plus bash or a code-running sandbox only if skills execute scripts.",
      "Progressive disclosure in practice: only name, description and path go into the system prompt, never the skill body; the agent picks a skill and then reads its instructions. Demo skills were company-research (web search → markdown), report-HTML and report-PDF (markdown → output).",
      "Descriptions are routing signals: the PDF skill only fired when its description said 'use the skill only when user ask for a PDF report' — 'PDF' is the trigger word. Descriptions should be aligned to the user's request rather than describe the skill, be distinct enough to avoid confusion, and not go stale; those are the reasons skills fail to trigger.",
      "In agentic products (unlike coding harnesses) skills are mostly model-driven, not user-invoked, because non-technical users shouldn't carry the cognitive load of remembering which skill to name.",
      "Cut skills by user intent, not by data model: he started with narrow data-shaped skills (estimation analysis, fundamentals) and had to refactor multiple times into intent-shaped ones — an 'earnings preparation' skill instead of estimate analysis, a 'pre-market briefing' skill instead of news-and-analyst-rating skills.",
      "A model upgrade broke their agent with zero changes to any skill file: the new model weighted the beginning of the skill heavily while their critical instruction sat at the end. Hence: skills are contracts versioned to a model, and evals must be rerun on every model upgrade.",
      "Routing scales in stages: system-prompt stuffing works for a few skills; past ~10, shortlist with embeddings/similarity search or a smaller shortlisting model; at hundreds of skills you need a hierarchy of skills, metadata filters and governance.",
      "Five aspects of skill-library governance, each borrowed from code practice: admission (an automated gate on the registry with human in the loop, analogous to PR review — should this skill exist or fold into an existing one), ownership (named skill owners, maintained by application teams, like CODEOWNERS), lifecycle (semantic versioning, deprecation warnings, change logs), coherence (periodic audits and skill validation checks), and boundaries (allow-listed, access-controlled tools per skill)."
    ],
    "takeaways": [
      "If you're building an agentic product rather than a coding agent, add skill support to your own harness: a registry of name/description/path, concatenate those into the system prompt, give the agent a file read tool, and add bash or a sandbox only if skills run scripts.",
      "Write skill descriptions as routing signals phrased in the user's language and containing the trigger words users actually say ('PDF report'), keep them mutually distinct, and refresh them — most non-triggering skills are description problems.",
      "Name and split skills by user intent ('earnings preparation', 'pre-market briefing'), not by your data model, and expect to refactor the skill library repeatedly as real use cases arrive.",
      "Build evals for skills and rerun them on every model upgrade — treat a skill as a contract versioned to a model, not documentation; also put critical instructions near the top, since models differ in where they attend.",
      "Plan the routing mechanism to change, not just tune, as the library grows: prompt-stuffing under ~10 skills, embedding/small-model shortlisting beyond that, hierarchy plus metadata filters plus governance at hundreds."
    ],
    "topics": [
      "agents",
      "skills",
      "agent-harness",
      "evals",
      "routing",
      "governance",
      "enterprise",
      "fintech"
    ],
    "tools": [
      "Anthropic",
      "Agent Skills / SKILL.md",
      "Claude Code",
      "Codex",
      "FactSet",
      "NVIDIA (NVDA, demo query)",
      "embeddings / similarity search",
      "X (Twitter)"
    ],
    "quotes": [
      {
        "text": "the role of engineer is shifting from shipping features to shipping harnesses. Harnesses that are smooth vehicles for your skills to run.",
        "at": "04:47",
        "url": "https://www.youtube.com/watch?v=7jjudsEhBtM&t=287s"
      },
      {
        "text": "it's very important to run evals and skills without evals are really just wishful thinking.",
        "at": "12:17",
        "url": "https://www.youtube.com/watch?v=7jjudsEhBtM&t=737s"
      },
      {
        "text": "Skills are not the documentation and a lot of people treat them like that and skills are really the contracts versioned to a model.",
        "at": "12:24",
        "url": "https://www.youtube.com/watch?v=7jjudsEhBtM&t=744s"
      },
      {
        "text": "routing mechanism doesn't get tuned as you scale. It changes the mechanism itself",
        "at": "16:46",
        "url": "https://www.youtube.com/watch?v=7jjudsEhBtM&t=1006s"
      }
    ],
    "words": 2516
  },
  "summary_url": "/#7jjudsEhBtM",
  "transcript": {
    "html": "/transcripts/7jjudsEhBtM.html",
    "txt": "/transcripts/7jjudsEhBtM.txt",
    "vtt": "/transcripts/7jjudsEhBtM.vtt"
  }
}