A gradual starting sequence

  1. Choose an established coding tool, such as Claude Code, Codex, or GitHub Copilot. Pocock suggests his skills as an optional starting point and advises against spending too long choosing tools. 36:01 36:19
  2. Learn how to stop the agent before giving it extensive freedom. His driving analogy is to learn the brake first, avoiding unwanted work and wasted rate limits. 36:43
  3. Open an empty folder, optionally use VS Code to inspect files, and start with a concrete idea you want to build. A personally interesting result strengthens motivation and the learning feedback loop. 37:09 37:27
  4. Watch the agent closely and learn what it and its surrounding tool actually do. Build trust through observed behavior. 37:54
  5. Once you understand its behavior, expand to a second agent, custom skills, or connections to other data sources. 38:09

What earning trust looks like

  • Pocock calls this climbing the trust ladder, a phrase he attributes to Lauren Tan. Initial distrust is appropriate; trust should grow as you learn the system. 36:19 36:43
  • His own integration example is dictating a task into his watch, passing it into a system and to-do list, then having an agent triage it. 38:28
  • Herk's bicycle analogy adds that supervision continues after you increase the number of agents. 39:00 39:24

A disagreement about ambitious tasks

  • Herk asks about public calls to give models your most ambitious goals. Pocock recommends the opposite starting point: simple, boring tasks with outputs you can inspect and measure, such as drafting replies. 39:44 40:04
  • Brainstorming and Grill Me can help develop ideas, but Pocock's main benefit is automating routine work so he can focus on strategy. 40:23 40:39
  • His example is generating first drafts of written content and then iterating on them. He recommends tasks whose mistakes are easy to correct, and says this made his own work more enjoyable. 40:39 41:03