This episode explains slash goal, a new AI interaction primitive in Codex and Claude Code that shifts from turn-based prompts to a continuous, self-evaluating loop. It details how to define clear success criteria and evidence for tasks, covers best practices for coding and knowledge work, and highlights why this approach increases autonomy while keeping the user in control.
Summarized by Podsumo
Slash goal transforms AI interaction from turn-based to a continuous loop where the model self-evaluates and iterates until completion criteria are met.
It's most effective for tasks with durable objectives, uncertain paths, and strong, inspectable evidence (tests, citations, artifacts).
Non-coding knowledge work like literature reviews, market landscapes, and claim audits can benefit, especially when user-provided rubrics define success.
The feature offers user control via pause/resume/clear commands, and the thread dynamically accumulates context as the goal progresses.
Adoption is growing: Claude Code adopted the same '/goal' primitive, recognizing it as a new industry standard for agentic autonomy.
"Don't tell it what to do, give it success criteria, and watch it go."
"The skill that wins is engineering the intent: strategic context and how success will be measured, so the agent can make better autonomous decisions."
"Slash goal might be the most consequential thing we've shipped in Codex. The value of good instructions has never been higher."