GPT-6 Astra · OpenAI's guidance, decoded

Five changes to get more out of Astra.

OpenAI published how to prompt GPT-6 Astra and clean up your skills and AGENTS.md. Here's what to steal.

Some of the instructions you wrote for older models are now making Astra worse. OpenAI says so in their own words. This page takes their guidance, shows you the exact lines, and turns each one into something you can paste into Codex today.

Five swaps. Each one traced to OpenAI's wording.
5
Changes, each traced
to OpenAI's own line
6
Copy-and-paste prompts
(OpenAI's, verbatim)
1
Audit prompt for the
instructions you already have
2
Official OpenAI sources,
linked and quoted

Quotes are from OpenAI's developer blog post "Rethinking skills and prompts for GPT-6 Astra" and the GPT-6 Astra model guidance page. Everything else is our reading of them.

Free Guide

Get the Astra guide.

Drop your email and the whole guide unlocks right here, with copy buttons on every prompt and a Markdown download. You'll also get one practical AI system most weeks.

  • The five changes, with the exact OpenAI wording behind each one
  • OpenAI's own prompt snippets, verbatim, with copy buttons
  • The one-line versions you can drop into a task prompt
  • Where each instruction belongs: AGENTS.md, the task prompt, or a skill
  • The audit prompt that finds outdated rules, conflicts and skills that load when they shouldn't
  • The rest of the guide: skill descriptions, decision boundaries, persistence, delegation
No spam. Unsubscribe anytime. You'll join "Stay Ahead" by AI Systems Lab.
Start here

Your old instructions were written for a different model.

OpenAI's opening line says it plainly: "With more capable models, what used to require a lot of handholding and scaffolding no longer does." The rest of the guidance is about removing scaffolding, and adding a few precise lines in its place.

Your first move
Run the audit first. It tells you which of the five you actually need.
  1. Open Codex in the project you use most.
  2. Paste the audit prompt. Read what it finds before you change anything.
  3. Apply changes 1 to 4 where the audit says the instruction is missing or contradicted.
01 · The source

What OpenAI actually published.

Two documents. They overlap, and together they are what this page decodes.

Developer blog

Rethinking skills and prompts for GPT-6 Astra

The short essay about cleaning house: skill descriptions, progressive disclosure, AGENTS.md, decision boundaries and persistence. Ends with the suggestion to have Astra audit your setup.

Model guidance page

GPT-6 Astra prompting best practices

The reference page with the actual prompt snippets: initiative and follow-through, instruction following, writing style, subagent delegation, testing and verification.

OpenAI, model guidance page
"GPT-6 Astra is stronger at general instruction following than our previous models, giving you greater control over its behavior. It can be more sensitive to instructions contained in skills and other files, such as AGENTS.md. We strongly recommend auditing skills and other files accessible to your model for instructions that could influence its behavior."

That sensitivity is the whole story. Astra follows your instructions more literally than GPT-5.6 Sol did. So a rule you added to stop an older model going too far now stops Astra from finishing.

02 · Change one

Make it show the finished work before asking for approval.

You should be approving something you can see, not a plan.

Instead of
"Help me improve this page."
Say
"Build the changes, test them locally and show me the preview. Ask before publishing."
OpenAI, model guidance page
"Prompt the model to ask for approval only after preparing a concrete, reviewable result. This avoids blocking the task before the model has done the work it can, and often leads to quicker task completion."

OpenAI's snippet for this goes in your AGENTS.md or system instructions. It's the one that stops Astra from pausing at "shall I proceed?" before it has done anything.

OpenAI's snippet · approval after the work · AGENTS.md
Before asking the user clarifying questions, you should complete the work that is already authorized from context and necessary to make the proposed action concrete and reviewable. The user should be approving a concrete, reviewable result. For example, before deploying a change, writing to an external application, merging a PR or publishing a site, do all the required work first so that user approval is the final step. You don't need user permission for reversible tasks, read-only actions, reviews or fixes, or anything for which authorization is provided earlier in the session or strongly implied from the task instruction. Do not introduce unsolicited warnings, disclaimers, approval flows, or safety/compliance checklists due to hypothetical risk.

Notice the last line. Publishing still needs your approval. Everything reversible before it doesn't. That is the split the whole guide keeps coming back to.

There is a companion snippet for the "can you" problem, where Astra answers "yes I can" and stops:

OpenAI's snippet · "can you" means do it · AGENTS.md
When the user's prompt indicates a request for action, such as "can you...", "I want to...", "help me..." and similar expressions, treat these as instructions to do the work and take action. Do not stop at acknowledging capability (e.g. "Yes…"), proposing a plan, or offering to continue. Do not settle for a partial or "helpful enough" solution that does not fully satisfy the user's task to save time, effort or tokens. If a task requires sustained work, complete all the necessary work until the intended outcome is fulfilled.
03 · Change two

When it gets stuck, ask which instruction stopped it.

Fix the conflicting rule instead of typing "keep going" for the tenth time.

Instead of
"Keep going."
Say
"Which SKILL.md made you stop? Name the file and quote the rule."
OpenAI, model guidance page
"Unclear or conflicting guidance in a skill file may cause the model to pause and block work early. Make the priority of user instructions and skills explicit."
"Asking the model to identify the skill and instruction that caused it to pause or change direction can also be effective in providing transparency into model behavior."

Two lines from OpenAI do this. The first sets the pecking order. The second makes Astra show its working when it stops.

OpenAI's snippet · user beats skill · AGENTS.md
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.
OpenAI's snippet · name the file, quote the rule · AGENTS.md
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.

Why it works: OpenAI says to use this "to find silent and conflicting guidance when your application loads many skills and instruction files such as AGENTS.md." Once Astra quotes the rule, you delete or rewrite the rule. The stall doesn't come back.

04 · Change three

Tell it what good writing sounds like.

Left alone, Astra formats everything. Lists, tables, headers, a summary at the end.

Instead of
"Write it up."
Say
"Lead with the answer. Plain English, short paragraphs, concrete examples."
OpenAI, model guidance page
"The model tends toward detailed, formatted responses and may use recurring phrases across sessions. Specify the writing style and structure your application needs."
OpenAI's snippet · prose over formatting · AGENTS.md or task prompt
Default to using clear, concise paragraphs, each developing one main idea. Use lists only when the information is genuinely parallel, sequential, or easier to compare, and avoid nested lists unless the hierarchy cannot be expressed clearly in prose. Use plain, simple language: familiar words, concrete examples, and precise verbs. Prefer active voice and direct statements. Make sure to state the main point clearly and early, then develop it with the explanation and detail the reader needs. Let each sentence build on what came before. Develop the points that matter and provide enough support to be useful.

OpenAI also published a list of the words and patterns it wants gone. Worth pasting as-is if Astra writes anything a human reads.

OpenAI's snippet · the slop list · AGENTS.md or task prompt
Avoid using slop words or phrases like "Bottom Line:" in conclusions, "delve," "foster," "leverage," "it's worth noting," "importantly," "Question? Answer." or "This isn't about X. It's about Y.", "genuinely" or hyphenated compound descriptions and adjectives. Do not use concluding summary statements such as "In short:..", "The simplest mental model is:...". State the intended action directly. Avoid adding what you won't do, what will remain unchanged, or how you'll separate or categorize results. Do not use contrastive framing such as "X, not Y" or "X—not Y" that introduces an unprompted alternative that the user didn't ask about. Avoid invented compound labels like "exact-head checks" and "editorial-row layouts", vague qualifiers, and canned transitions; use plain verbs and prepositions to state the actual relationship directly.

For technical work, OpenAI adds one more line: "Use plain language over jargon, and reference technical details only to the degree that it helps illustrate an idea or your work to the user."

05 · Change four

Match the checks to the job.

Changing a button label shouldn't trigger endless rounds of testing.

Instead of
"Always run the full test suite before you finish."
Say
"Run the checks relevant to this change. Repeat them only if something fails or you make another change."
OpenAI, developer blog
"Previous models needed encouragement to run tests and check their work. GPT-6 Astra does that on its own, so the same instructions can lead to unnecessary testing."
OpenAI's snippet · testing calibrated to the change · AGENTS.md
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.

The flip side: if there is a test suite you know is safe, tell Astra it's allowed to run it without checking in. OpenAI's example for an AGENTS.md line:

OpenAI's example · permission for a safe workflow · AGENTS.md
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.

Delete the old line. If your AGENTS.md still says "always run all tests" from the GPT-5 days, that rule is now doing the opposite of what you wanted.

06 · Change five

Get it to audit your old instructions.

Before you download another massive prompt pack, get the model to check the instructions you already have.

OpenAI, developer blog (closing line)
"A new model is a good opportunity to clean your house, but you don't need to review everything manually: ask GPT-6 Astra to do an audit based on what was discussed in this article, then go build something you wouldn't have attempted before!"

OpenAI doesn't publish the audit prompt itself. This is ours, built from the checklist in their article. Paste it in the root of the project you use most.

Our prompt · audit my instructions · paste into Codex
Audit every instruction file you can read in this project: AGENTS.md at every level, every SKILL.md and its supporting docs, and any other context files. Use OpenAI's guidance in "Rethinking skills and prompts for GPT-6 Astra" (https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra) and the GPT-6 Astra model guidance page (https://developers.openai.com/api/docs/guides/latest-model?model=gpt-6-astra) as the standard. Report, then suggest changes. Do not edit anything until I approve. Flag: 1. Outdated rules written for older models: instructions that push you to run tests you would run anyway, read a stack of docs before every edit, or stop for review after a first implementation. 2. Conflicting instructions: any two rules that disagree, and any skill guidance that could make you pause, ask for permission, or leave work unfinished. Quote both sides. 3. Skills that load when they don't need to: descriptions that are long, broad, or overlap with another skill. For each, propose a shorter description that names the specific workflow it is for. 4. Skills written as step-by-step recipes where a goal and constraints would do. 5. Boundary language that is stronger than the risk: "never", "always ask", "do not proceed without" on reversible or read-only actions. 6. Missing lines from OpenAI's guidance: user instructions take precedence over skills; name and quote the skill that stopped you; approval only after a concrete, reviewable result; testing calibrated to the change. For each finding: the file, the exact line, why it's a problem under Astra, and the proposed replacement. Group by file. Keep the report in plain paragraphs and short lists, no tables.

What you should get back: a file-by-file list of lines to delete, lines to shorten, and lines to add, with nothing changed yet. Approve the ones you agree with. Then let it edit.

Also from OpenAI: Codex has an official $openai-docs skill that can "apply the recommended changes in this guide" for API migrations. Our audit prompt is for your instructions, which that skill doesn't cover.

07 · Placement

Where each line belongs.

The guidance names three homes for instructions: skills, AGENTS.md and the task prompt. Put each change where it will apply at the right moment, and nowhere else.

ChangePut it inBecause
1 Approval after the workAGENTS.mdYou want it every session. The one-line version ("show me the preview, ask before publishing") goes in the task prompt when a task ends in a publish.
2 Name the file, quote the ruleAGENTS.mdStalls happen anywhere. You want the explanation without asking for it.
3 Writing styleAGENTS.md for repos that produce copy; task prompt otherwiseA coding repo doesn't need prose rules on every task. A content repo does.
4 Testing calibrated to the changeAGENTS.mdReplace the old "always run all tests" line rather than adding next to it.
5 The auditTask prompt, once per projectIt's a job, not a standing rule. Re-run it when you switch models.

Keep AGENTS.md short. OpenAI: "Because AGENTS.md applies whenever the model works in your repository, you should frequently revisit each instruction and ask yourself whether it's still needed."

08 · The rest of the guide

Four more things OpenAI said that didn't fit in 60 seconds.

Skill descriptions are too long.

OpenAI: "many descriptions are far too long, and when you add too many skills, Codex starts shortening their descriptions to fit. The model ends up seeing less of each description, making it harder to know which skill to pick."

OpenAI's "bad" example
Create and validate Postgres schema migrations. Use when working with databases, queries, models, or persistence.
OpenAI's "good" example
Create and validate Postgres schema migrations. Use when adding or changing a migration, or reviewing its rollout.

Their rule for multi-workflow skills: "make the root document a minimal router that points to supporting docs and scripts." And a warning about recipes: "overly specific guidance can now hinder results where it previously helped."

Stop making it read everything.

OpenAI: "Requiring a stack of docs or a full repo map before every edit is excessive for a typo fix. GPT-6 Astra can work out what it needs to read without being pushed to review the whole project before every change."

OpenAI's "bad" example
Before every edit, read architecture.md, database.md, and deployment.md.
OpenAI's "good" example
Use architecture.md for service boundaries, database.md for schema changes, and deployment.md when preparing a deployment.

Soften the boundaries you wrote for older models.

OpenAI: "If you stated boundaries previously because you wanted to prevent other models from going too far and you're now switching to GPT-6 Astra, consider updating that language: Astra could take it too seriously and may stop work where you'd actually be happy for it to continue."

Define done before you start.

OpenAI: "GPT-6 Astra can feel more tentative about when to stop. It may reach a first implementation and come back for your review while there's still work to do." Their fix: "If the task includes getting the implementation running, inspecting the result, and fixing what fails, make that part of the request."

OpenAI's snippet · bias towards action · AGENTS.md
You should infer the user's intent and task scope from the instructions and prior conversation context. Your job is to bias towards action and carry the user's intended task to completion. When the user expresses intent to perform new work or fix an existing issue, persist until the user's intended goal is complete. Progress autonomously towards the user's goal (e.g. creating isolated worktrees / checkouts if needed, resolving merge conflicts, read-only actions, creating draft PRs etc.) unless they are clearly destructive or irreversible.

If you run subagents: OpenAI also notes Astra "may delegate less often than desired" and gives a one-line prompt for it: "If at any point you can parallelize work by delegating tasks to another agent (no matter if you are the root or subagent), you should do so using collaboration tools if it could save time or improve quality."

Troubleshooting

If it still misbehaves.

ProblemWhat to try
Still asks "shall I proceed?" before doing anythingCheck for an older "always ask before" rule in AGENTS.md or a skill. Use change 2 to make Astra quote it, then delete it. OpenAI's own troubleshooting note points at the initiative snippet in change 1.
Stops after the first implementationPut "run it, inspect the result, fix what fails" in the task itself. OpenAI: "A requirement to stop for review after the first implementation will pull the model toward an earlier stopping point."
Runs a huge test pass for a tiny changeFind and remove the old "always run all tests" line. Add the change 4 snippet in its place, not beside it.
Answers in tables and bullet lists when you wanted prosePaste the prose snippet from change 3 into the task prompt. If it persists across sessions, move it to AGENTS.md.
Loads a skill that has nothing to do with the taskThe description is too broad. Rewrite it to name the specific workflow, the way OpenAI's migration example does.
The audit finds nothingAsk it to list every instruction file it read. If it skipped nested AGENTS.md files or skill sub-docs, point it at them and run again.
Keep it

Take the guide with you.

GPT-6 Astra: five changesMarkdown file. All five changes, every OpenAI snippet, the audit prompt and the placement table. Drop it next to your AGENTS.md.

Official sources.

OpenAI updates these pages. If a snippet here differs from the live page, use the live page.

Want the full library of systems like this?

This is one build. AI Systems Lab is where the rest live: practical AI systems you can run in your own business.