Get More from GPT-6 Astra: Simplify AGENTS.md and Skills

You ask for a one-line copy change. The agent reads a stack of project documents, writes a plan, and asks whether it should continue. When that happens, check your AGENTS.md: does it still contain a rule beginning with “Before every change…”?

The model has changed. Instructions added to prevent mistakes months ago may no longer fit the work you want it to do.

On September 11, OpenAI's Eric Provencher published Rethinking skills and prompts for GPT-6 Astra. The post recommends revisiting skill scope, persistent instructions, and task completion criteria to reduce outdated constraints and irrelevant context.

Codex can help with that review. Asking it to identify a few specific changes is a more useful starting point than asking it to “optimize all my settings.”

Start with this audit prompt

Open Codex in the project you want to review and confirm that you are using Astra. If you still need to check model access, start with the GPT-6 Astra in Codex guide.

Use this OpenAI article as a reference:https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra
Review the AGENTS.md files, overrides, and skills that applyto this project. Find instructions that cause unnecessary work,conflicting behavior, or avoidable pauses.
First identify the active instruction sources and skill descriptions.Read skill bodies and references only when needed to verify a finding.
Look for:- Mandatory reading unrelated to the task.- Small changes expanding into repository-wide or repeated checks.- Overbroad skill triggers or conflicting instructions.- Repeated confirmation within authorized work, or premature stopping.
For each finding, give the file location, original wording,concrete effect, and smallest useful change.Separate explicit official guidance from project-specific inference.Preserve valid engineering constraints and authorization boundaries.Explain anything you cannot verify.Return suggestions only; do not edit files or invent findingsto reach a target count.

This is an audit template written for this article, not an official OpenAI prompt. Pasting it produces suggestions you can review; it does not automatically rewrite your setup.

Check whether the response names real files, specific sentences, and situations that trigger them. A paraphrase of every rule, or an even longer replacement handbook, has not yet solved the problem.

Know which instructions are always present

AGENTS.md and skills enter a task differently.

The Codex AGENTS.md documentation describes an instruction chain combining global guidance and project directories. AGENTS.override.md can change which file is selected. A clear root-level file does not establish that no other instructions apply to your current task.

Skills initially expose their names and descriptions; the complete SKILL.md is read after selection. The official skills documentation also describes a size budget for that initial list. With many skills, descriptions may be shortened and some entries may be omitted. A description that tries to cover every kind of work makes selection harder.

Consider a weekly-report skill:

• “Use for tasks, files, communication, and teamwork” could match almost anything.
• “Turn existing work records into a weekly report; use when drafting, editing, or reviewing that report” gives it a recognizable job.

The name and description help the agent choose the right entry point. Report formatting, unusual fields, and script arguments can live in the material needed when that work begins.

Change one sentence and watch what it changes

The examples below show how to connect an edit to observable behavior. They are illustrative, not universal repository settings.

Attach reading requirements to a task

Suppose a project says: “Read all product, architecture, and release documents before every change.”

That rule applies even when the entire job is changing a button label. A narrower instruction could be: “For button copy changes, inspect the surrounding page and copy guidelines. Consult release instructions when preparing a release.”

On the next comparable task, you should be able to observe the difference: the agent finds the page and relevant guidance, then starts editing. If it still reads every document first, find the instruction responsible for that work.

Give verification a finish line

“Keep checking until there are no problems” is difficult to accept as a completion criterion. It specifies neither the checks nor a stopping point.

For a form error-message change, ask the agent to reproduce the error state, inspect the wording and any overlap, and run the checks required by the affected module. Once those pass, it can deliver the result; new failures give it a reason to continue.

Verification remains part of the task, with a visible completion condition. The wording to reconsider is the open-ended requirement to repeat it. Payment flows, permissions, and data consistency still need checks proportionate to their actual risks.

Say exactly where confirmation belongs

One file says “finish the task before returning.” Another says “stop for confirmation after every step.” Continuing and stopping each violate an instruction.

If you want local edits, fixes for affected test failures, and a page inspection, include them in the authorization for the task. Identify decisions reserved for a responsible person, such as a production deployment, separately. Steps already covered by the authorization can proceed together; new scope or risk gives the agent a concrete reason to pause and explain.

An audit can identify the conflict. Decisions about team authority belong to the people responsible for them; an agent should not remove a rule simply because the rule is inconvenient.

Check the task you just sent, too

Even with a tidy configuration, “take a look at how to fix this” can reasonably produce an explanation and stop there.

For a small page bug, a task could read:

Fix the login form's obscured error message.Make the code change, inspect normal, error, and submitting states,run the relevant checks, and fix problems introduced by the change.When those conditions are satisfied, summarize the changeand the verification results.This task ends after local verification.

That request identifies the problem, the expected result, and the endpoint. Replace the checks with the ones your project actually needs. Adding “also comprehensively optimize the project” would make the scope vague again.

If other models use the same repository instructions, keep a few representative tasks for comparison. A prompt that seems redundant for Astra may still help another agent used by the team.

Find out whether the cleanup helped

Choose a task you perform regularly. Keep the code baseline, model, reasoning effort, and tools the same, then run it with the old and revised instructions. Check result quality first. Compare file reads, tool calls, confirmation pauses, total time, and measured usage. When runs vary, observe several examples before drawing a conclusion.

Skipping two irrelevant files can also reduce context carried forward. Removing a repeated check can avoid another round of tool calls. The actual savings depend on the original instructions and the task. API cost calculations also need separate input, cached-input, and output figures; the coding-agent API cost guide explains those categories.

The instructions most worth keeping often contain information the model cannot reliably infer: which check the project actually uses, which directory needs special care, and whose authorization a particular action requires. Make room for those facts and define a clear finish line. The next task can then show whether the edits helped.

Based on OpenAI's September 11, 2026 post and the official AGENTS.md and skills documentation, checked September 15, 2026. The audit prompt and worked examples were written by Agent.Space.