Keep pulling the thread on Jason Liu.
Thibault from the Codex team described the "/goal" feature as potentially the most consequential feature shipped in Codex.
Andrej Karpathy of Anthropic has stated that LLMs are exceptionally good at looping to meet specific goals when provided with success criteria rather than explicit instructions.
Claude Code released a feature equivalent to Codex's "/goal" primitive and also named it "/goal", indicating a move toward a shared industry standard.
Blitzy claims its agents can autonomously deliver over 80% of the work required for a software project after ingesting a codebase.
According to OpenAI's documentation, a strong "/goal" definition includes six components: outcome, verification surface, constraints, boundaries, iteration policy, and a block stop condition.
The "Codex Maxing" concept, based on a blog post by OpenAI's Jason Liu, is a set of tips and best practices for using OpenAI's Codex.
Jason Liu's blog post on "Codex Maxing" outlines approximately nine techniques for using Codex effectively for coding and other knowledge work.
Pavel Huren explained that with the "/goal" feature, a user states the desired outcome, and the model then loops, self-evaluates, and stops when the task is complete.
Nicholas Bustamante of Microsoft described the "/goal" process as an initializer agent turning user intent into a plan.md file, worker agents making progress, and a judge agent deciding on completion.
Sean Wang (Swyx) characterized the evolution of AI interaction as moving from "/skill" (preset prompts) to "/plan" (human-refined inputs) to "/goal" (AI-evaluated outputs), representing an increase in autonomy.
The "/goal" primitive is fundamentally different from a standard prompt, acting as a "finish line contract" that defines what should be true, how success is checked, and what must remain intact.
The "/goal" feature in Codex includes user commands such as "/goal pause", "/goal resume", and "/goal clear" to allow for intervention without losing progress.