Skip to content

Codex custom instructions experiment

Experimental: this is a local Codex model-catalog experiment, not installed Agent OS behavior.

The experiment replaced a 17,730-character shared GPT-5.6 instruction prompt with:

  • a 4,889-character shared prompt for Sol, Terra, and Luna;
  • a 1,541-character read-only prompt for Auto Review.

The design keeps request-derived authority, external/destructive boundaries, preservation of user work, proportional verification, and fresh evidence before completion. It removes duplicate warnings and reduces speculative edge-case analysis, unnecessary approval questions, over-validation, and routine status narration.

The controlled smoke suite completed 31 runs across Sol, Terra, Luna, and Auto Review. The candidate preserved the tested authority and review behavior. A first live desktop task after installation produced two meaningful updates, identified two evidence-backed PR blockers, avoided hypothetical risk lists, and stopped at a concrete next gate. That observation is anecdotal; the full behavioral matrix remains open.

The complete experiment record includes:

  • both exact prompts;
  • the nine-operation JSON Patch contract;
  • baseline and candidate measurements;
  • installation hashes and startup proof;
  • rollback instructions;
  • limitations and the first live observation.

Read the canonical experiment record on GitHub.

OpenAI references used for the design:

A personal framework, published in the open.