Auditing and trimming Codex workspace instructions

# I audited the instructions running my Codex workspaces

After reading Eric Provencher's Rethinking skills and prompts for GPT-6 Astra, I asked Codex to audit my own setup. I wanted to know which instructions were still useful and which ones I could get rid of. (https://x.com/pvncher) (https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra)

I started with the configuration and global instructions, then expanded the audit to the installed skills. Before applying the recommendations, I backed up the files so I could undo the changes.

## Cleaning up 143 skills

I learned that Codex reads skill names and descriptions to decide which instructions to load. Eric recommends keeping those descriptions short and specific, and moving detailed procedures into files Codex can read when it needs them.

I had 143 local skill files, including six system skills that I left intact. I cut the other skills' descriptions from 56,560 to 12,392 characters, a 78.1% reduction.

As an example, I had five skills for reporting Power Platform issues. Those became one skill that asks which product the issue affects. I also retired an old Canvas app alias and moved skill-authoring advice into documentation.

In my Power Platform collection, add-data-source handles Canvas apps and add-datasource handles code apps. The names differ by one hyphen, but they work with different types of apps.

For the code-app integrations, I put connection lookup and build checks in one shared procedure. SharePoint, Dataverse, Excel, and Teams still have their own instructions, but I can now update the common steps in one place.

I ended up with 136 local skill files. I also made sure that the retired skills are stored outside the folder Codex searches, so I can bring them back if needed.

## Choosing when a skill runs

I made five skills, including marketing planning and a review helper, run only when I ask for them. This setting goes in the skill's agents/openai.yaml file:

I also broke up several long skill files. The authentication skill was nearly 3,000 lines. Its main file now points to separate instructions for the work being done, so diagnosing a login problem doesn't require loading all the instructions for setting up or replacing authentication.

Some supporting files instructed Codex to log out, regenerate authentication code, or request broader SharePoint permissions than the operation needed. I changed them to preserve existing authentication and request the permissions needed for the operation.

## Stop asking me to approve routine decisions

My test-driven development skill told Codex to ask me to approve which parts of the app the tests should exercise. If I ask for a bug fix, I expect it to read the existing tests and work that out. I revised the skill to use the requested behavior, existing tests, and affected code to choose what to test.

A shorter prompt for the same approach is:

I kept the rules protecting credentials and unrelated files, checking which account Codex uses, and requiring approval for new production dependencies. I also kept the instruction to inspect unfamiliar tests before running them. 

I also updated the code-review instructions, I changed what Codex reviews before committing a fix, and the review now includes uncommitted edits and new files.

One review script reported a clean review whenever the review command exited successfully. We reproduced it with a test command that printed a high-priority finding and then exited with code zero.

I changed the script to report that the command finished and require assessment of its output before calling the review clean. The shortened shell example looks like this:

The script now also checks the repository's default branch instead of assuming main, and stops if a required fetch fails. We checked these changes with tests in disposable repositories.

## Keeping the work going

I removed the old context-guard settings based on a 260,000-token threshold and disabled the Ponytail skill. The recovery instructions now tell Codex to record its progress at useful stopping points and continue from that record.

This is the information I want carried forward:

## Try it on your own setup

All 136 remaining local skill files passed the skill validator. I also checked that Codex could load the updated skills and that the retired ones no longer appeared.

I also audited my Mac mini, which had its own configuration and installed packages. I prepared a separate audit prompt for that machine. You can adapt this shorter version:

If you’ve been adding instructions to Codex every time something goes wrong, or when you want to expand capabilities I’d give those files another read. 

I found rules I no longer wanted it to follow, along with useful guidance buried in skills that needed cleaning up. @pvncher's article gave me a reason to go back through them.

Our agentic AI tools grow and improve, and as they do, we need to allow them to molt away their old skin. Revisit, refactor, and prune your configs and workflows regularly.