You ask GPT-6 Astra to change one button on your site. It is much smarter than GPT-5, but it stops to ask you something after every step, and the answer comes back laid out like a report. Reading OpenAI's model guidance, these turn out to be tendencies it was designed with, and the fixes sit on the same page.
Below are the 8 items I put together after going through the original text. Each has three parts: what you run into, how OpenAI says to fix it, and one line you can paste straight into your instructions. After that come the two other official points (splitting work across subagents, not over-testing), then one combined prompt you can copy as a block.
Source: the GPT-6 Astra section of OpenAI's Model guidance page (developers.openai.com/api/docs/guides/latest-model), compiled from the official text, 2026-09-06.
I have pasted the whole block into my own setup. The one I felt most was number 4: have it scan for rules that fight each other first, and it stops switching direction halfway through.
Last updated: 2026-09-06
Fix 1: it keeps asking you questions
What you run into
Astra asks as soon as anything is fuzzy, going back and forth several times before it starts. OpenAI says this is by design: when more information would change the result, it leans toward asking first.
The official fix
Tell it to bias toward action, infer the scope from your instruction and the context, and carry what the user wanted through to done, asking only when it would truly go the wrong way.
Paste this into it
Guess what I mean from the context. If you can guess it, just do it, and carry the thing I want through to done. Ask me only when you would otherwise go in the wrong direction.
Fix 2: it stops halfway for approval and adds its own warnings
What you run into
It likes to ask "is this okay" before doing anything, and adds warnings, disclaimers, and safety checklists you never asked for.
The official fix
Produce something visible first, then hand it to you for review. No warnings or approval steps that were not asked for.
Paste this into it
Produce something I can look at first, then ask me what to adjust. Do not add warnings, disclaimers, or safety checklists I did not ask for.
Fix 3: it treats "help me" as a conversation
What you run into
"I want to...", "help me...", "could you..." get read as discussion, so it replies with what it could do, or with a plan only.
The official fix
Spell out that these phrasings mean do the work, and that it should not stop at describing what it can do, proposing a plan, or finishing half of it.
Paste this into it
When you see sentences like "I want", "help me", or "could you", always treat them as a request to do the work. Do it, then report back. Do not hand me a plan only.
Fix 4: long instructions fight each other and it picks one
What you run into
Astra reads skills, AGENTS.md, and other config files. As soon as the rules in them fight each other, it easily stops, or switches direction halfway through.
The official fix
Say plainly that your instruction outranks the rules inside skills, and have it scan every rule it can currently read for places where they contradict each other.
Paste this into it
First check whether any current instructions or files contradict each other, and list them. When there is a conflict, my message here wins over any skill or config file.
Fix 5: it stalls without saying what stalled it
What you run into
When it stops, you get one vague reason and no way to tell which file caused it.
The official fix
When it stops out of nowhere, ask which file it read and which line blocked it, and have it paste that text back.
Paste this into it
Why did you stop just now? Name the file you read, paste the original text of that instruction, and explain how you understood it.
Fix 6: answers come out looking like reports
What you run into
Left alone it gives you headings, tables, and bullets all at once, and repeats the same phrases across different conversations.
The official fix
Short paragraphs, one point each. Lists only when the items are really parallel, ordered, or easier to compare, and never nested. Familiar words, concrete examples, clear verbs, say who did what.
Paste this into it
Answer in short paragraphs, one point per paragraph, in everyday words with clear verbs. Use a list only when the items are really parallel, and never nest lists.
Fix 7: it reads as AI-written at a glance
What you run into
delve, foster, leverage, it's worth noting, importantly, Bottom line: these words are where the AI flavor comes from.
The official fix
Ban these stock phrases. Also tell it not to invent a Y just so it can write "X rather than Y", since that shape reads as AI. Use a verb and say what you are doing.
Paste this into it
Do not use delve, foster, leverage, it's worth noting, importantly, or Bottom line. Do not write contrast sentences shaped like "X rather than Y". Say what you are doing.
Fix 8: too much jargon piled on
What you run into
In technical docs it packs in every term of art, and ordinary readers cannot follow.
The official fix
Plain words first. Technical detail only when it helps the reader understand.
Paste this into it
Write so an ordinary reader can follow. Use a term of art only when there is no plainer way to say it, and explain it the first time it appears.
Two more that matter if you code with Astra
It can split work out, but rarely does so on its own
Astra can hand work to subagents in parallel, but by default it delegates less than you want. OpenAI suggests writing it straight into your instructions: whenever work can be split out to another agent in parallel and that saves time or improves quality, do it, and keep messages between agents readable for a human, with spaces between words and numbers.
Whenever work can be split out to another agent in parallel and that saves time or improves quality, do it. Keep messages between agents readable for a human.
Do not test a small change to death
For reversible, low-impact changes it writes piles of tests that just mirror the implementation. OpenAI suggests: tests should be meaningful and actually needed. Run the tests that match the change, and once they pass, keep going to done. Widen testing only for new changes, failures, or concerns that are still open.
For reversible, low-impact changes, do not write tests that just mirror the implementation. Run tests proportional to the change, and once they pass, carry on to finish the task. Widen testing only when a new failure or concern shows up.
Combined version: paste one block
If you do not want to paste them one at a time, use the whole block below. Put it in your custom instructions, or at the top of AGENTS.md for Codex.
Guess what I mean from the context. If you can guess it, just do it, and carry the thing I want through to done, and ask me only when you would otherwise go in the wrong direction. When you see sentences like "I want", "help me", or "could you", always treat them as a request to do the work. Do it, then report back. Produce something I can look at first, then ask me what to adjust. Do not add warnings, disclaimers, or safety checklists I did not ask for. When instructions contradict each other, my message here wins over any skill or config file. When you stop, name the file and the line that stopped you, and paste the original text. Answer in short paragraphs, one point per paragraph, in everyday words with clear verbs. Use a list only when the items are really parallel, and never nest lists. Do not use delve, foster, leverage, it's worth noting, importantly, or Bottom line. Do not write contrast sentences shaped like "X rather than Y". Use a term of art only when there is no plainer way to say it, and explain it the first time it appears.
Common questions
Why does GPT-6 Astra keep asking questions?
OpenAI says it leans toward asking first when one more question would change the result, and that this is by design. The fix is item 1.
Where do these instructions go?
Custom instructions, or the first message of a conversation. For Codex, write them into AGENTS.md.
Do they work on Claude or Gemini too?
Writing style, banned stock phrases, and instruction priority carry over. The constant questions and stopping halfway are specific to Astra.
Of the 8, the two that make the biggest difference are items 1 and 4. With those written in, Astra goes from asking at every step to reporting once a whole chunk is done. You can try just those two first.
Want to turn AI into a system that actually works for you? See my services, or start with a free 30-minute chat.
This page is a free resource I made for you
I'm Leo Huang (kanisleo328), and I have built several products with AI. If this page helped, say hi on IG. I'll keep adding hands-on Claude Code guides.