GPT-6 Astra acts instead of asking, follows long instructions more closely, and formats everything unless you stop it. Here are the 7 changes, with a filled example for each one.
GPT-6 Astra landed in paid ChatGPT accounts this month, and it behaves differently from what came before. It acts instead of asking. It follows long instructions much more closely. It formats everything unless you tell it not to.
So the habits you built to work around an older model are now working against you. Here are the 7 changes, in the order I'd make them. The first is a setting. The other 6 are one line each in how you write the ask.
Where these came from. Six of them are in OpenAI's own prompting guidance for this model, which is public. The effort call is Ruben Hassid's, from a week of using it every day, and I've marked it as his.
Each one is a change to how you ask, not a new tool to learn. Read the change, take the example, and put it in your next real prompt.
The change: before you type anything, pick how hard it thinks. Light is fast and shallow. Medium balances quality against waiting. High is for hard reasoning and deep planning.
Ruben Hassid keeps his on Medium after using Astra every day for a week. He calls it the best mix of speed, cost and thinking. He also works only in Work or Codex mode. That part is his setup from real use.
OpenAI does not publish a best setting. Their own guidance calls effort a tuning knob and says task clarity matters more than the level you pick. Which is what the other 6 changes on this page are.
The change: start the prompt with what you want to be holding when it stops, not with what you want it to look at.
Give me a ranked list of my [5 best months], tell me how you grouped them, and flag anything based on too little data. Deliver the ranked list first, then your reasoning.
The change: when you attach something, give it a job. A file is evidence, a style example, a template, or instructions. Astra guesses if you don't say, and it guesses worst on style examples.
Use the [brand guide] for voice and tone only, not for facts. Use the [product brief] for all product details. If they disagree about a feature, flag the conflict before you write about it.
The change: Astra pauses to check far more than older models did. That old opener now interrupts work you wanted finished. Name only the decisions you actually want a say in.
Ask me only about [the audience and the deadline]. Make reasonable assumptions on everything else, complete the work, and list your assumptions at the end.
The change: left alone it writes long, formats everything, and reuses the same stock phrases. Describe the shape you want once and put it in your custom instructions so you never ask again.
Default to clear, concise paragraphs, each developing one main idea. Use lists only when the information is genuinely parallel, sequential, or easier to compare. Keep paragraphs to two or three sentences. Explain an unfamiliar term the first time it appears. Start with the useful finding and end with one practical action.
This one belongs in your custom instructions, not in every prompt. Set it once and every answer after it comes back in your shape.
The change: you can send a new instruction mid-job and Astra folds it into the plan instead of you cancelling and rewriting. OpenAI calls this mid-turn steering.
Use the March column instead of the totals while it is still going.The change: Astra tests and verifies thoroughly by default, which is slow on small reversible work. Tell it what is worth checking instead of letting it check everything.
Before you deliver, confirm the totals match the source, identify duplicate entries, and check that every date falls inside [the reporting period]. Report any mismatch you could not resolve. Do not run broader checks than that on this task.
These are in the same OpenAI document and are aimed at people running longer or more complex work. If the 7 above are enough for your week, stop there.
can you or I want to is read as an instruction to do the thing, not a question about whether it could.OpenAI's own guidance says to treat the effort setting as a tuning knob and not the main way to get better answers. Task clarity, constraints, and saying what the output should look like matter more than the level you pick.
Which means changes 2 through 7 on this page are worth more than change 1. Start with the finish line.