Edessa

Run your SOP inside Codex

Last updated: October 3, 2026

Codex reads a file before it does anything, so pointing it at a process is one file.

You put the file in the right place and it gets read.

What can Codex read?

Codex reads AGENTS.md files: one in your Codex home directory for things that apply everywhere, and one per directory along the path it is working in. It reads every AGENTS.md from the project root down to the folder you start it in, so a workflow file next to the work applies to that work.

Two details worth knowing. There is a size budget, 32 KiB by default across the files it collects, so a long workflow can crowd out everything else. And empty files are skipped, which is a quiet way for a broken export to look like no export.

The details move, so check OpenAI’s AGENTS.md guide rather than a blog post. The install steps were checked against the vendor docs on October 3, 2026.

Where does the file go?

  1. Export the workflow as Markdown.
  2. Save it as AGENTS.md in the directory the work happens in, or in your Codex home directory if it applies to everything: AGENTS.md
  3. If one already exists, add the workflow to it rather than replacing it.
  4. Start Codex in that directory. Ask for the task; the file is already read. If it seems to be ignoring it, check the size budget before you rewrite the process.
  5. Or export it as a Skill and unzip it into ~/.agents/skills, where Codex picks it up when a task matches its description: unzip -o ~/Downloads/<skill>.zip -d ~/.agents/skills/
Try Edessa free

Three asks worth trying

Once the process is loaded, the useful prompts are not “summarise this.” They are the ones that use the fact that Codex can now see how your business actually works.

  • "Follow the process in AGENTS.md and stop after the step that needs my approval." Ask it to stop there. An agent that runs the whole way through is useful once and risky after that.
  • "Which steps in this process are you least sure how to do?" Where the agent hesitates points to where the write-up may be thin, which makes this a free documentation review.
  • "Trim this process to the steps that actually change the output." A written process can carry steps that exist only because somebody once did them.

How do you keep it current?

This is where documentation goes stale. The process changes, the file does not, and six months later the skill is confidently wrong. Keeping it right by hand means editing a Markdown file every time a step changes.

The alternative is re-recording. Do the task once the new way, replace the file with the new export, and the thing your AI tool reads matches what you actually do. That takes about as long as the task does.

The recording method is here, including the manual version if you would rather not use a tool for it.

Who wrote this. Thaddeus Demeke, Founder of Elevaite and the founder of Edessa. Five years at Genentech, most recently as a Senior Data Scientist, then building AI systems inside small businesses across agencies, home services, real estate and ecommerce. I run my own processes this way.

See how the AI SOP generator works

Related reading

© 2026 ElevaiteWorks LLC. A product of Elevaite.