Audit Your AGENTS.md Before GPT-6 Astra: Remove Dead Rules That Now Mislead

OpenAI's newly released GPT-6 Astra model excels at software engineering, browser automation, and multi-step professional tasks across code, web pages, and applications. Crucially, it follows instructions more faithfully and adjusts mid-task when requirements change. Because of this heightened sensitivity, the official documentation now explicitly recommends auditing AGENTS.md and Skills files — vague or conflicting rules can cause premature stops or unnecessary verification loops.

Why Clean Up?

Rules accumulate over time to prevent errors, but each rule should be traceable to a specific problem it prevents. The author's own writing workspace illustrates the issue: AGENTS.md still required calling mp-writer, yet the corresponding entry point no longer existed — the skill had been replaced but the signpost remained. Simply switching models does not fix stale configuration.

An important detail: Skills load only their name and description initially; the full body is read on demand. Therefore, auditing must focus on over-broad trigger conditions, duplicate processes, and expired requirements — not on total file size as a proxy for context savings.

How to Clean Up

The author packaged the audit workflow into a reusable skill named audit-agents-md. Install it in the project terminal (requires Python 3):

npx skills add https://github.com/lltx/skills --skill audit-agents-md

Then invoke it in a Codex conversation:

$audit-agents-md 审计当前项目内的所有技能和 AGENTS.md，只输出问题与改写建议，暂不修改文件。

The tool inventories all files, then checks each rule for: trigger conditions that capture ordinary requests, minor edits forced into heavy processes, and referenced paths that no longer exist. Every finding includes file location, trigger scenario, consequence, and a rewrite suggestion. Rules that were only scanned — not fully read — are also flagged for manual review.

After reviewing the report, remove unused Skills only after confirming no other workflow references them; same-named directories may be symlinks to the same file. Finally, run routine tasks to verify that shortcuts are restored and original requirements still hold. The goal is not merely shorter files, but removing expired rules that still command the agent — while preserving business boundaries and hard-won lessons learned.