{
  "video": {
    "id": "DouK81O8R3o",
    "title": "One Day, $1,000 in Tokens: An AI War Room for Legacy Code — Ford Prior, GEICO",
    "duration": 1123,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "DouK81O8R3o",
    "title": "One Day, $1,000 in Tokens: An AI War Room for Legacy Code — Ford Prior, GEICO",
    "one_liner": "GEICO's Ford Prior bet $1,000 of AI tokens and one day on a multiplayer 'AI war room': key decision-makers in one Slack channel, one human driver, and FleetKit, a set of markdown rules steering Claude Code agents. It mapped a legacy claims domain end to end and delivered a change by a 4pm demo, so the 'outer loop' of approvals and decisions kept up with AI-accelerated coding.",
    "summary": "Prior argues that AI has 'basically obliterated' the build step of the SDLC. At a risk-management company like GEICO, though, major (V1→V2) changes are still held back by the outer loop: getting the right people in a room to make decisions, plus the intelligence to back those decisions. His answer is an AI-native war room. It has a roster of decision-makers and subject matter experts, one 'driver' running the AI tools, a comms channel, design tools and code repos, all wired to agents through FleetKit, a markdown fleet manager built with Anthropic and forked from Anthropic's code modernization plugin. GEICO ran it three times across its rentals (uplift), damages (transform) and salvage (reimagine) domains, in legacy claims systems with millions of lines of code.",
    "key_points": [
      "GEICO found that after patches and minor upgrades, engineers had only about 25% of their bandwidth, 'on a good day or less', left for major changes. So they invested in GEICO-managed AI agents, on several harnesses and models, to handle patch and minor work from spec to deployment.",
      "The bet: with $1,000 of tokens in one day, map one domain end to end (including graph representations of the business topology and the complex legacy code), have the agent find and propose one serious change (an agent-callable API into the legacy claim systems), and have it deployed to their skills agent platform for a 4pm demo to his boss.",
      "Three runs, using the uplift/transform/reimagine taxonomy from Anthropic's code modernization plugin. Rentals: extended an existing API, and the agent found a way to save 8,500 calls per month by giving customers up-to-date rental information. Damages: built new agent-callable APIs to unlock a larger workflow. Salvage: produced a fully vetted migration plan with API specs and design mock-ups, green-lit by product and tech at every level, and found a prod bug.",
      "Split of ownership: the vendor supplied the replaceable parts (harness, model, FleetKit, code modernization plugin). GEICO supplied what can't be replaced: context (knowledge base, software standards, 'past decision archaeology', tool lists), methods (skills and domain-specific ways of building software) and judgment (who can decide what at which gate). These live in repos that both agents and humans can add to and correct.",
      "FleetKit is a structured set of rules, protocols and instructions written in markdown, optimized for the Claude harness (specifically Fable) and for architecture transformation. It takes a goal, breaks it down, dispatches leaders to research and choose an approach, and puts decisions in a human-in-the-loop queue (e.g. GitHub issues). Once approvals are in, it dispatches builders, who use beads for tracking, to produce a 'material change'. It isn't open source yet.",
      "Workflow: a Claude Code session with the plugin adopts or creates a mission Slack channel, invites the war room members, and follows a nine-step process. At milestones it pings the specific owner from the roster and produces HTML worksheets (Claude artifacts) that people fill in, which save automatically. It also produces design documents and architecture visualizations that humans edit in tools like draw.io or Figma and the agent then rereads. The plugin deterministically checks that each step produced its required artifacts.",
      "Two kinds of agents: cloud 'code dives' that crunch and graph the legacy domain-specific languages, and local agents on the driver's laptop (the runtime) with access to his MCPs, including enterprise ones not yet in production, described to the agent by a bindings file.",
      "Lessons: prompt caching was 'way higher than I thought it would be', subagent behaviour was quirky, and benchmarks against recently written code as golden data found Sonnet 5.5 wrote the legacy code as well as Fable or better: 'model selection for the win.'"
    ],
    "takeaways": [
      "Fork the code modernization plugin's steps and adapt them to your own process instead of using it as-is. Put the human-approval worksheets early, where most of the human input is needed.",
      "Encode your irreplaceable context, methods and judgment (who decides what at which gate) in versioned repos that both agents and humans can correct. That deep context is built once and reused.",
      "Never let Claude running Fable set up crons to check on work: each check launches a new Fable session and burns the budget.",
      "Benchmark cheaper models against your own recent code as golden data before defaulting to the most expensive model.",
      "Find a group in your organization that struggles to make decisions, then deploy a war room there, drive it and control its cost. On the roadmap for GEICO: stronger gate enforcement, better context search than plain markdown files, and audio transcription as a data source."
    ],
    "topics": [
      "agents",
      "legacy modernization",
      "human-in-the-loop",
      "sdlc",
      "multi-agent orchestration",
      "cost management",
      "enterprise ai"
    ],
    "tools": [
      "GEICO",
      "Anthropic",
      "Claude Code",
      "Fable",
      "Sonnet 5.5",
      "FleetKit",
      "Anthropic code modernization plugin",
      "beads",
      "GitHub issues",
      "Slack",
      "Claude artifacts",
      "MCP",
      "draw.io",
      "Figma",
      "graphify",
      "Obsidian"
    ],
    "quotes": [
      {
        "text": "developing software looks less like playing in a canvas and more like working in a factory.",
        "at": "11:23",
        "url": "https://www.youtube.com/watch?v=DouK81O8R3o&t=683s"
      },
      {
        "text": "Wow, that's actually a great question that we've never thought about identifying an edge case either in the business or with the technology.",
        "at": "13:42",
        "url": "https://www.youtube.com/watch?v=DouK81O8R3o&t=822s"
      },
      {
        "text": "do not let uh Claude running Fable set up crons to check on work uh because you'll it'll just launch a Claude uh a Fable session each time and burn your budget immediately.",
        "at": "16:07",
        "url": "https://www.youtube.com/watch?v=DouK81O8R3o&t=967s"
      },
      {
        "text": "y'all find some group in your organization that doesn't know how to make decisions, deploy this, drive it, control the cost, make it yours.",
        "at": "17:57",
        "url": "https://www.youtube.com/watch?v=DouK81O8R3o&t=1077s"
      }
    ],
    "words": 4137
  },
  "summary_url": "/#DouK81O8R3o",
  "transcript": {
    "html": "/transcripts/DouK81O8R3o.html",
    "txt": "/transcripts/DouK81O8R3o.txt",
    "vtt": "/transcripts/DouK81O8R3o.vtt"
  }
}