Keep pulling the thread on Jason Liu.
The episode charts the progression of human-AI interaction from simple, preset prompts (`/skill`) to human-refined plans (`/plan`), and now to the autonomous, outcome-driven `/goal` primitive. This shift signifies a move up the abstraction stack, where users define the 'what' and 'why', leaving the 'how' to the AI agent.
The core mechanism of the `/goal` feature is its ability to loop, act, and self-evaluate against a defined success criterion without continuous human intervention. This concept, highlighted by figures like Andrej Karpathy, allows the AI to navigate uncertain paths and persist on a task over an extended period.
Successfully using `/goal` requires a different skill set than traditional prompt engineering. It's about meticulously defining the 'finish line contract'—the final state, how to verify it, and the rules of engagement—rather than just crafting an initial instruction.
The episode explores how the `/goal` primitive, born from software development, can be adapted for general knowledge work. The key is to reframe tasks like literature reviews, due diligence, or market analysis as processes that produce an 'audit'—a verifiable artifact with clear evidence, citations, and confidence levels.
The rapid adoption of the `/goal` command, including its name, by a major competitor (Claude Code) suggests the emergence of a standardized set of powerful interaction primitives across the AI industry. This mirrors the development of standardized commands in other areas of computing.