The /goal Trick: Make Codex or Claude Code Loop Until a Goal Is Verified

Key Takeaways

• The move: /goal + a goal in Codex or Claude Code. The agent loops and keeps working until it verifies the condition is hit, instead of stopping after one pass.
• The goal must be verifiable. Give it a binary, checkable condition — e.g. "edit this blog post until it is under 500 words." That's a clear yes/no the agent can confirm on its own.
• Turn subjective goals into objective ones. "Make this funnier" is too subjective — so build the test into the goal: "set up three AI judges; not complete until all three judge this funnier than the original." The loop continues until the eval passes.
• This is what 2026 AI looks like. Miller frames it as loops, systems, and AI-as-a-judge — the real skill is figuring out how to make outcomes verifiable and how to set up evals that decide when something is actually good.
• It's a two-second test. Open Codex or Claude Code, type /goal and a verifiable target, and watch it loop on itself and do the work.

Commands & Code Mentioned

# Inside Codex or Claude Code:
/goal edit this blog post until it is under 500 words        # verifiable: clear yes/no

# For a subjective target, build the eval into the goal:
/goal make this blog post funnier — set up three AI judges; not complete
      until all three judge it funnier than the original