How to trim skills, AGENTS.md, and task prompts for GPT-6 Astra

If you have been running Codex for a year, your repo is probably carrying skills, AGENTS.md rules, and task prompts written for models that needed more scaffolding. OpenAI’s Eric Provencher argues that with GPT-6 Astra those leftovers matter more than they did on earlier releases. Astra is more instruction-sensitive: a bloated skill catalog wastes context, and guardrails written for Sol or Luna can stop work early.

This is a rewrite pass. Use OpenAI’s bad-versus-good pairs, then define “done” so the model does not stall.

Narrow skill descriptions to when-to-use triggers

Skills are prompts stored as Markdown, sometimes packaged with resources and bundled scripts. They are most useful for a specific workflow or when using certain apps. Each skill’s name and description load into context so Codex knows when to use it.

Many descriptions are far too long. Add too many skills and Codex starts shortening descriptions to fit. The model sees less of each trigger and picks worse. Descriptions can also contradict each other or over-emphasize when a skill should load, so the model pulls in instructions that do not help the task.

Keep descriptions as short as possible while making the trigger obvious.

Bad

Create and validate Postgres schema migrations. Use when working with databases, queries, models, or persistence.

Good

Create and validate Postgres schema migrations. Use when adding or changing a migration, or reviewing its rollout.

The bad line can fire anytime the model touches a database, query, model, or persistence. The good line fires when you are adding or changing a migration, or reviewing its rollout.

A common way to create skills is the $skill-creator skill. OpenAI recently updated its guidance to address these failure modes.

Prefer a thin router over a giant recipe

Reading a skill takes context, brings you closer to compaction, and injects guidance that may not apply. For skills with multiple workflows, make the root document a minimal router that points to supporting docs and scripts. Give the model enough to know where to look without forcing a full read.

Many skills were written as elaborate itineraries. Models now handle nuance and ambiguity better, so overly specific recipes can hinder results where they once helped.

Repository skills also guide other contributors’ agents, which may use different models. Guidance that helps Sol or Luna may overconstrain GPT-6 Astra. Write for the models that will load the file.

Replace blanket AGENTS.md reads with task-conditional pointers

Because AGENTS.md applies whenever the model works in the repository, revisit each instruction and ask whether it is still needed. Requiring a stack of docs or a full repo map before every edit is excessive for a typo fix. Astra can work out what it needs to read without reviewing the whole project first.

Bad

Before every edit, read architecture.md, database.md, and deployment.md.

Good

Use architecture.md for service boundaries, database.md for schema changes, and deployment.md when preparing a deployment.

Prompting a pre-read of those files on every edit burns context and slows work. Contextual pointers still help. Keep the docs updated too.

Let Astra test; grant permission for the safe suite

Previous models needed encouragement to run tests. GPT-6 Astra does that on its own, so the same instructions can lead to unnecessary testing. The latest-model prompting guide notes the same habit: for coding tasks, Astra tends to be thorough before considering a task complete, and on smaller tasks that can mean broader tests than the change requires.

If you want to calibrate that, OpenAI publishes this prompt:

Do not write tests for reversible, low-impact changes that mirror the implementation. If you do choose to verify your work with tests, make sure that the tests are meaningful and necessary to verify implementation.

Run tests appropriate to the change and complete required checks. Once those pass, broaden or repeat testing only when new changes, failures, or unresolved concerns justify it; otherwise, continue toward completing the task.

Astra is thorough, but it can be tentative about how far to take a task. You can use AGENTS.md to give permission for a workflow you know is safe, such as a local test suite:

The local tests use disposable fixtures and have no production access. Run them, fix failures caused by the requested change, and rerun affected tests without asking for approval at each step.

Loosen ask-first fences; make user instructions win

Pay attention to how you describe boundaries. If a previous model did things without permission, you may have added strong language to make it ask first. That can still be useful, but GPT-6 Astra is OpenAI’s most aligned model, with better judgment, and it will not perform tasks unless it knows they are safe. If you wrote those fences to restrain other models and you are switching to Astra, update the language. Astra can take old restrictions too seriously and stop where you would be happy for it to continue.

The latest-model guide adds that Astra is stronger at instruction following and more sensitive to skills and files such as AGENTS.md. Unclear or conflicting skill guidance may cause it to pause and block work early. Make precedence explicit:

The user's instructions take precedence over guidelines provided in a skill. If explicit user instructions conflict with a skill's instructions, prioritize the user's instructions.

If a skill still derails the run, ask for the file and the line:

If a skill causes you to ask for permission or confirmation, pause, leave requested work unfinished, or diverge from the user's intent, name and link to the exact SKILL.md file you read, quote the relevant instruction, and briefly explain how it applies. Distinguish explicit skill requirements from your interpretation of guidelines.

Spell completion in the task prompt

If you are used to GPT-5.6 Sol taking a request and continuing for long stretches, Astra can feel more tentative about when to stop. It may reach a first implementation and come back for review while there is still work to do.

Define completion before starting. If the task includes getting the implementation running, inspecting the result, and fixing what fails, make that part of the request. A requirement to stop for review after the first implementation pulls the model toward an earlier stopping point—keep it only if you actually need that gate. If you want it to keep exploring beyond a first pass, say what you want explored and where it should stop.

That matches Astra’s collaborator bias in the latest-model guide: it is more likely to ask when extra input could materially change the result. Prompt for follow-through when the user’s wording already implies authorization, and ask for approval only after a concrete, reviewable result.

You do not have to review every file by hand. Ask GPT-6 Astra to audit skills, AGENTS.md, and task prompts against this guidance, then go build something you would not have attempted before.

Next step: Rewrite one high-traffic skill description to a when-to-use trigger, turn any multi-workflow skill into a thin router, replace blanket AGENTS.md reads with task-conditional pointers, and put run-inspect-fix into your next Codex task.