Park Long Jobs in Files. Keep Rules as a Map.
What Codex adds is not better chat. It can write agreements and progress into project files so the next shift can continue. Files hand off. Chat does not.
Short changes are enough with the five steps. Work that crosses hours and sessions adds one Codex-specific judgment: write agreements and progress into the repo so the next shift — human or agent — continues from files. Officially it reads project instructions before it works. Long-horizon practice calls durable project memory the most important technique.codex-agents-md codex-long-horizon
After reading, you should be able to answer:
- Why are project instructions a map, not an encyclopedia?
- Which three kinds of files does a long job need at minimum?
- What do humans still review, and what do they stop reviewing?
Keep rules short and close to the code
Project instructions hold standing agreements: how tests run, when it must ask, the bad pattern and the safe path. Keep them short. Put global habits on the machine, repo norms at the root, and service-specific rules next to that code — nearer files win. Officially, empty files are skipped, and a temporary swap uses an override file instead of deleting the base.codex-agents-md
Dump every sentence into one file and task context gets crowded out, every rule looks equally important, and you cannot tell whether any rule was followed. When OpenAI used Codex to build an internal product, the first giant instruction file failed immediately. They later compressed it to about a hundred lines that pointed into structured docs.openai-harness The in-site versions are the project contract and context layering.
Three files catch the next shift
A long job needs three kinds. Names can change; jobs cannot mix:
- Milestones: what each step must produce, which command accepts it, repair before moving on.
- Runbook: how to work now, and keep the diff small enough to read.
- Status log: what happened and why, readable after a few hours away.
That is the user-side move. It is not another lesson in building agents. The mechanism is in long tasks need a handoff and a terminal state.
Humans review the calls, not every line
People still set the target, the constraints, and the course corrections, and they review architecture and product choices. They do not watch every line. The direction of long work is less babysitting, more delegation with guardrails. Next: how this usage breaks.