Codex knowledge surfaces

July 31, 2026 ยท View on GitHub

Status: verified Last verified: 2026-07-31 Sources: official-codex-manual, karpathy-llm-wiki-gist

Current understanding

Codex benefits from several durable surfaces, but they serve different scopes. Use the prompt for one task, AGENTS.md for required repository conventions, a skill for a reusable procedure, an MCP connector for live external data, and automation for a proven recurring workflow. Local memories can help recall prior work but should not be the only copy of required team guidance.

A living wiki adds a separate evidence layer: it stores synthesized understanding, uncertainty, experiments, and promotion decisions without injecting the entire knowledge base into every task.

Evidence

  • OpenAI documents AGENTS.md, skills, plugins, MCP, memories, and scheduled tasks as separate customization surfaces.
  • Karpathy's pattern separates immutable sources, a compiled wiki, and the schema that teaches the agent how to maintain it.

Implications

Keep mandatory rules compact in AGENTS.md. Load wiki pages only for tasks that need their evidence, and promote stable procedures into focused skills rather than asking Codex to reread the full wiki.

Promotion status

The surface-selection guidance is already represented in modules 04, 05, and 11. This page remains an evidence map for maintainers.