Big Idea

The argument in one line.

GPT-6 Astra follows instructions more literally, asks more clarifying questions, and defaults to list-heavy prose, so the fix is feeding it real context and fewer vague rules, not more reasoning effort.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You're already using ChatGPT Work, Codex, or GPT-6 Pro for real work and want to stop wasting a limited Astra message allowance.
You carry over custom instructions and prompting habits from older OpenAI models and haven't adjusted them for Astra.
You want a documented, copy-pasteable set of prompting rules sourced directly from OpenAI's own guide.

SKIP IF…

You're not using GPT-6 Astra or ChatGPT Work/Codex — these specific adjustments are tuned to this model's behavior, not general prompting advice.
You want a broad introduction to prompt engineering rather than Astra-specific fixes.

TL;DR

The full version, fast.

GPT-6 Astra behaves differently from older OpenAI models: it follows instructions more literally, stops to ask clarifying questions more often, and defaults to long, list-heavy, repetitive prose. Pulled from OpenAI's own developer guide and the default instructions baked into Codex, the fix is six adjustments: hand over real files and define the result instead of raising reasoning effort, paste a style directive that forces plain paragraphs over lists, explicitly tell Astra to act instead of proposing a plan, start every task on the lowest thinking setting and raise it only when needed, define what 'done' means with one verification check, and strip old blanket rules like 'always ask me first' out of custom instructions since Astra now obeys them to the letter.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:26

01 · How OpenAI's engineers prompt GPT-6 Astra

Creator explains he researched X, Reddit, and OpenAI's own prompting guide to distill six tips, several of which invert what worked on older models.

00:26 – 00:54

02 · What Astra costs and the Plus limits

API pricing (~$10/M input, $50/M output) and ChatGPT Plus's 5-45 message cap per five hours, roughly half of GPT-5.6 Sol's allowance.

00:54 – 01:24

03 · Where to find Astra (ChatGPT Work and Codex)

On Plus, Astra only lives inside ChatGPT Work and Codex, not normal chat; Pro/Business users see it in normal chat as GPT-6 Pro.

01:24 – 02:03

04 · OpenAI's official Astra prompting guide

The six tips all trace to OpenAI's developer-site guide, whose lines are also baked word-for-word into Codex's shipped default instructions.

02:03 – 03:29

05 · 3 things that changed with Astra

OpenAI says Astra follows instructions more literally, stops to ask clarifying questions more often, and defaults to long, list-heavy, repetitive formatting.

03:29 – 05:05

06 · Tip 1: Give it the files and the context

Reasoning effort can't substitute for missing information; hand over real files, name the audience, and specify the output shape, per OpenAI's rewritten example.

05:05 – 05:37

07 · Keep your files in a ChatGPT project

Recurring jobs should live in a ChatGPT Project folder so every new chat inherits the shared files and instructions automatically.

05:37 – 06:12

08 · Free AI growth plan for your business

Sponsor-style plug for a free personalized AI growth plan from the creator's consultancy.

06:12 – 07:35

09 · Tip 2: Stop the lists and repeated phrases

Paste OpenAI's own style directive (plain paragraphs, lists only when genuinely parallel, no filler words) to stop Astra's default list-heavy, repetitive formatting.

07:35 – 08:00

10 · Put it in your custom instructions

Save the style directive once into ChatGPT or Codex custom instructions instead of repasting it every chat.

08:00 – 09:03

11 · Make it write in your own style

Turn on ChatGPT's writing-style personalization against connected email/Slack/Docs, and use a lighter model like Astra-lite for simple rewrites.

09:03 – 10:51

12 · Tip 3: Stop it from asking questions

Paste OpenAI's directive to treat 'can you' / 'help me' as instructions to act, paired with one boundary: require approval before sending or publishing anything.

10:51 – 11:37

13 · Which thinking setting to use

The light-to-high reasoning picker resets between model generations; OpenAI says start Astra at light and raise it only when needed.

11:37 – 12:31

14 · Astra on light vs GPT-5.6 Sol on high

OpenAI's benchmark shows Astra on low outscoring Sol on high; for simple jobs, skip Astra and use Sol or Luna instead.

12:31 – 14:28

15 · Tip 4: Tell it what done means

Define completion in three parts before starting: the result wanted, what must not change, and one check to confirm it's finished.

14:28 – 15:09

16 · Tip 5: "Think harder" does nothing

Typing 'think harder' doesn't change reasoning effort; only the thinking-level setting in the picker does.

15:09 – 17:03

17 · Tip 6: Clean up your old custom instructions

Old blanket rules get followed literally by Astra; have it read your instructions, flag every stop-causing line, and keep only the one or two that matter.

17:03 – 18:39

18 · All 6 tips on one real prompt

A messy real-world request gets rewritten in one pass via a meta-prompt applying all six Astra rules, producing a tight four-part brief.

18:39 – 19:23

19 · Turn it into a skill

Closing suggestion to save the rewritten meta-prompt as a reusable skill/template, plus sign-off and resource plug.

Atomic Insights

Lines worth screenshotting.

GPT-6 Astra costs about $10 per million input tokens and $50 per million output tokens on the API, several times pricier than the model it replaced.
On ChatGPT Plus, Astra is capped at roughly 5 to 45 messages every five hours, about half the allowance GPT-5.6 Sol gave.
Astra doesn't appear in ChatGPT Plus's normal chat window at all — it only lives inside ChatGPT Work and Codex, showing up in normal chat as 'GPT-6 Pro' only on Pro or Business plans.
OpenAI's own help page states plainly that increasing the reasoning effort cannot supply missing information or access.
Astra follows instructions so literally that old habits like spelling out every step or saying 'check with me first' now work against you, producing constant pauses.
On OpenAI's own coding benchmark, Astra on its lowest reasoning setting scores higher than GPT-5.6 Sol on its highest.
The person who runs Codex at OpenAI said it directly: if you were happy with Sol on high, move down to low or medium on Astra.
For simple jobs, OpenAI's own guidance is to skip Astra entirely and use Sol or Luna instead, since they stretch a usage allowance much further.
OpenAI defines 'done' as three parts: the result you want, what must not change, and one check to confirm the work is actually finished.
Piling on multiple verification checks doesn't make Astra more thorough — it just re-verifies the same work and burns extra usage.
Typing 'think harder' in a prompt does not change Astra's reasoning effort; only the thinking-level picker in the model selector does.
Old blanket custom instructions like 'always ask before doing anything' get followed to the letter by Astra, which can stall work it would otherwise finish.
The fix for legacy custom instructions is to have Astra itself read them, quote every line that would make it stop or ask permission, and keep only the one or two boundaries that actually matter.
Every prompting fix in the video traces back to lines OpenAI's own engineers wrote into Codex's shipped default instructions, not a guess from a social media thread.

Takeaway

Six Rules For Prompting A Model That Actually Listens

WHAT TO LEARN

GPT-6 Astra follows instructions literally and stops to ask more often, so the fix isn't more reasoning effort, it's feeding it real context and trimming old vague rules down to the one or two boundaries that matter.

02What Astra costs and the Plus limits

Astra runs about $10 per million input tokens and $50 per million output tokens on the API, several times pricier than the model before it.
On ChatGPT Plus you get roughly 5 to 45 Astra messages every five hours, about half of what GPT-5.6 Sol allowed, so each message should count.

03Where to find Astra (ChatGPT Work and Codex)

On ChatGPT Plus, Astra never appears in the normal chat window, it only lives inside ChatGPT Work and Codex.
Pro and Business plans do get Astra in normal chat, listed under the name GPT-6 Pro.

04OpenAI's official Astra prompting guide

Every tip in the video traces back to OpenAI's own developer-site guide, not a guessed thread on social media.
The same guide's lines are baked word-for-word into Codex's default instructions, which have shipped since Astra launched.

053 things that changed with Astra

Astra follows instructions more literally than older models, so step-by-step habits now get followed to the letter instead of loosely interpreted.
It's more likely to pause and ask a clarifying question instead of making an assumption and continuing.
It defaults to long, heavily formatted answers with lists and tables, and tends to repeat the same phrases across sessions.

06Tip 1: Give it the files and the context

Raising the reasoning effort cannot substitute for missing information or access, OpenAI says this outright in its own help documentation.
Replace a vague request like 'make me a presentation' with one that names the source files, the audience, the exact output shape, and what to flag as unsupported.
The underlying rule: give it the sources, the templates, the constraints, and the checks that define a useful result.

07Keep your files in a ChatGPT project

For recurring jobs, put the files, chats, and instructions inside one ChatGPT Project folder instead of re-attaching them every time.
Every new chat started inside that project folder automatically inherits the shared files and context.

09Tip 2: Stop the lists and repeated phrases

Astra defaults to lists, tables, and repeated phrasing across chats unless told otherwise.
Paste OpenAI's own style directive: default to plain paragraphs, use lists only for genuinely parallel or sequential information, and avoid nested lists and filler words.
These exact lines already ship inside Codex's default instructions, which is why Codex output reads more like prose.

10Put it in your custom instructions

Instead of pasting the style directive into every chat, save it once inside ChatGPT or Codex custom instructions so it applies everywhere automatically.

11Make it write in your own style

ChatGPT's writing-style personalization can learn your actual voice once you connect email, Slack, Docs, or Notion.
Pick a lighter model like Astra-lite for simple rewrite jobs so you're not spending a limited Astra message on easy work.

12Tip 3: Stop it from asking questions

Phrases like 'can you,' 'I want to,' or 'help me' now risk getting answered with a plan or an offer to continue instead of finished work.
Paste OpenAI's directive: treat those phrases as an instruction to do the work, and don't stop at acknowledging capability or proposing a plan.
Pair that instruction with one explicit boundary, require approval before sending, publishing, or changing anything other people depend on.

13Which thinking setting to use

The light-to-high reasoning picker resets with each new model generation, so a setting that was 'high' on an older model isn't equivalent on Astra.
OpenAI's own guidance is to start every Astra task at the lowest setting and only raise it when the task genuinely needs more planning or checking.

14Astra on light vs GPT-5.6 Sol on high

On OpenAI's own coding benchmark, Astra on its lowest effort setting scores higher than GPT-5.6 Sol on its highest.
If Sol on high was your default, OpenAI's Codex lead says to move down to low or medium on Astra and expect the same or better results.
For genuinely simple jobs, skip Astra entirely and use Sol or Luna, since a limited usage allowance stretches much further on those models.

15Tip 4: Tell it what done means

Astra can hand back a first version while real work is still left, because it can be tentative about when to stop.
Define completion in three parts before the task starts: the result you want, what must not change, and how to check it's actually finished.
Keep it to one meaningful check, Astra already verifies its own work, so stacking multiple checks just reruns the same verification and burns usage.

16Tip 5: "Think harder" does nothing

Typing 'think harder' or 'think deeply' into a prompt does not change Astra's reasoning effort.
Only the thinking-level setting in the model picker actually changes how hard the model reasons before answering.

17Tip 6: Clean up your old custom instructions

Old blanket rules like 'always ask me before you do anything' or 'give me at least three options' get followed to the letter by Astra, often stalling work it would otherwise finish.
Have Astra read your existing custom instructions and quote back every line that would make it stop, ask permission, or leave work unfinished, with a suggested edit for each.
Keep only the one or two boundaries that genuinely matter and delete the rest, fewer rules, not more, is the fix.

18All 6 tips on one real prompt

A messy, realistic request can be rewritten in one pass by pasting a meta-prompt that applies OpenAI's own Astra rules automatically.
The rewritten version should name the result up front, the exact files and applications to use, what 'done' looks like, and end with a single check.

Glossary

Terms worth knowing.

GPT-6 Astra: OpenAI's flagship model as of its 2026 release, available inside ChatGPT Work, Codex, and, on Pro/Business plans, ChatGPT's normal chat under the name GPT-6 Pro.
ChatGPT Work: The part of ChatGPT built for larger jobs involving files and connected applications; on Plus plans it's the only normal-chat surface where Astra appears.
Codex: OpenAI's coding-focused product, built on the same model family, shipped with default instructions that already encode several of Astra's recommended prompting rules.
Thinking setting (reasoning effort): The light-to-high scale in ChatGPT's model picker controlling how much the model reasons before answering. It resets with each new model generation and does not respond to phrases like 'think harder' typed into the prompt.
Custom instructions: A persistent block of text in ChatGPT or Codex settings applied automatically to every new chat, used to lock in formatting rules, voice, and behavioral boundaries without repeating them each time.
ChatGPT Project: A folder inside ChatGPT that bundles chats, files, and instructions together so every new conversation started inside it automatically inherits that shared context.

Resources

Things they pointed at.

01:24linkOpenAI's official GPT-6 Astra prompting guide

01:46linkOpenAI Codex default instructions (GitHub commit)

05:37linkReprise AI free AI business growth plan ↗

00:00productSystems to Scale free Skool course ↗

Quotables

Lines you could clip.

03:45

“Increasing the reasoning effort cannot supply missing information or access.”

direct OpenAI quote, counterintuitive for anyone who just cranks up reasoning effort→ TikTok hook↗ Tweet quote

12:03

“If you're happy with Sol on high, move down to low or medium on Astra.”

specific, actionable, quotes OpenAI's own Codex lead→ IG reel cold open↗ Tweet quote

14:22

“Asking ChatGPT to think harder or think deeply will not switch it to a different thinking level.”

myth-busting one-liner anyone can verify immediately→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

So GPT -6 Astra has been out for a few weeks now and I've been trying to figure out how to prompt this better and how OpenAI's own engineers actually prompt this. So I've been going through X, going through Reddit, reading what OpenAI's team has been posting and reading the full guide OpenAI put out on how to actually prompt Astra.

And out of all of that, I pulled six different things. They're simple, you can change them today and a couple of them are just the opposite of what actually worked on the older models. And before I get into the tips, there's one thing that you need to know and that's Astra is...

not a cheap model now on the api it's about 10 per million tokens going in and 50 per million coming out and instead of chat gpt if you are on the plus version opening eye's own estimate is somewhere between 5 and 45 astro messages every five hours now the model before this GPT 5 .6 sole, that gets you about double that.

So every Astra message that you are going to be sending, of course you want it to be counting. And the other thing that I've seen trip some people up is that on Plus, you will not find Astra in the normal chat at all. It's going to be living inside of ChatGPT work, which is a part of ChatGPT that does all of the bigger jobs, the stuff with your files and all of your connected applications, and inside of Codex, which is just the coding side of things, obviously.

And if you're on Pro or Business, then yes, it will be showing up in the normal chat as well under the name GPT 6 Pro. Okay, so here I am inside of ChatGP to work. I can just open up the model picker and there of course is Astra and all the other preceding models.

Okay, so this here is OpenAI's guide and I wanna start here because every one of the six tips, it comes out of what OpenAI says is different about this particular model. So this guide, it lives on OpenAI's developer site, but the same guide, it also sits inside of OpenAI's own codex code on. GitHub and the commit that put it there came from one of OpenAI's engineers.

So this isn't some thread on X where somebody just guessed and it's slopped. These are the lines that OpenAI wrote for its own model. And it actually goes further than that.

So if you open up the Codex code, the default instructions that Astra runs with, they have several of these exact lines inside of them word for word. And they've been in there since the day that Astra launched. So inside of Codex, OpenAI does this for you.

Inside of ChaiGPT, you just have to add the lines yourself. Okay, now what does OpenAI say is different now? Well, there's three...

things number one it follows instructions more closely than the older models did so what they actually say is gpt6 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 the second one is it stops to ask more often so it's more likely to take a pause and ask you a question instead of just making an assumption and getting out of it and then number three it likes the long formatted answers and it tends to repeat these same phrases across their chats so what they say is the models they tend toward detailed formatted responses and may use recurring phrases across sessions specify the writing style and structure that your applications need now once you see those three you can see why the old habits actually start working against you and if you used to write out every single step asher will now follow every single step according to what you say if you used to say check with me first it's going to check with you quite often in the highest thinking setting

is no longer where OpenAI says to start. Okay, so jumping into all of the tips, by the way, we will have all of them completely free to download and available inside of my school community. They can be down below in the description if you wanna get access to that.

And of course, on how to actually grow your AI business to over $10 ,000 a month. So all of it's gonna be completely free and you'll be able to follow along very easily from that school community post. But anyways.

tip number one this is the easy one and it's about what you hand it because astra it can only work with what is actually going to be in front of you so if a file is missing or maybe a detail is missing it has two different options it is either going to be guessing or it will stop and it will ask you. And just turning the thinking setting up, it will not fix that natively.

In OpenAI's help page, it actually says in one line, increasing the reasoning effort cannot supply missing information or access. So the fix, it's actually kind of boring. You give it the files and you tell it who the result is actually going to be for.

You tell it what shape that you want back in return. So a great example of this, instead of saying, make me a presentation about our customer research, you instead will say something like this. Review the attached interview notes in the survey results.

Create an eight -slide presentation for the product leadership meeting. Focus on the three most common customer problems. Include supporting evidence, separate findings from recommendations, and flag any claims that are not well supported.

Use Google Drive for the source docs. Return a draft for my review before treating it as final. So that's OpenAI's own example that they actually gave.

And if you look at the difference, the first one is... make me presentation about our customer research the second one says which files to be reading who the deck is actually going to be for how many slides what you should be focusing on and that it should flag anything that is not backed by the evidence and interestingly enough openia has one line for astra on their model page that actually sums up the whole thing you give it the sources the templates the constraints and checks that define a useful result and for you if you do the same kind of job every single week you don't want to attach the same files every single time of course so you are going to be putting them inside of a project and a project this is just a folder inside of chat gpt that just keeps the chats the files and the instructions all neatly together and every chat that you start inside that folder you can effectively be using all of it so you can see one of my projects here it has main or it's called main and i pretty much post everything inside of here it's where all of my agents live it has all the information about me all the logins to all my different websites all that stuff

quick before we get into the second tip if you are a business owner and you are looking to use ai inside of your business in 2026 but you don't know how or where to start then make sure to check out the link down below in the description that is exactly why we created our ai roadmap we will create a completely custom growth plan for you and your business.

It's completely personalized. You answer a few quick questions and we will send that over to you. It is completely free.

It is completely custom. It takes about 30 seconds to actually fill it out. So if you're a business owner and you want to find out how and where to start with AI, then make sure to check that out.

Again, it's completely free. Okay, so tip number two. So once it has the right files, the next thing is actually how it's going to be writing.

And this one is specific to Astra. Now OpenAI, they do say this themselves. Astra, it tends...

towards detailed formatted answers lists and tables everywhere and it can use the same phrases over and over across different chats so opening i wrote the exact words to switch that off so in their case instead of saying write a follow -up email to a client whose invoice is two weeks overdue instead we can just paste this in 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 and then it goes on in length about avoid using slop words or phrases like so and so so the instead of that we have up here this is just a normal request a follow -up email about an Overdue invoice.

And OpenAI says that is exactly the kind of request that Astra likes to answer with lists and long headings. But the paste this, this is OpenAI's own block, and these two blocks are some of the lines that ship inside of Astra's defaults in Codex, where it's clear it is short paragraphs, one idea each, and it lists only when the information is actually a list.

So plain words, have active voice, and then they have their list of words that it should never actually use down below. Now for you specifically watching this video, you are not going to be pasting this into every single chat you actually just go into your settings and you're just going to be putting it into your custom instructions so if you just search that or you can navigate it to yourself you will see that we have our little codex section right here and we just throw this into our codex custom instructions anything that you actually enter it's going to be saved pretty much everywhere now additionally to that there's actually one thing that opening i cannot write for you which is your own voice so if you find yourself in the settings and we typically would just go to personalization you'll find at the bottom we have this writing section right here that we can enable so this is going to allow you to reference your writing style now all that really means is that once you connect your email your documents turn this on and giraffes they start sounding like you and then under opening eyes list you add your own you know never say these words because they're going to

slot once you actually enable it it's going to find all the different messaging so maybe slack or for some of your documents it's going to be notion and google docs or for email we use gmail so let's go ahead and use our writing style and then it's going to automatically enter this prompts to find a recent email or thread and draft a response using my writing style beyond this i'm actually going to say be pretty rigorous and grab a bunch of different emails and maybe a hundred different slack messages of mine i'm actually going to change this to gpt6 astralite because we don't need to use a crazy effort model for something this easy okay moving on to number three this is the complaint that i see most about astra so far and that's it just stops and asks you questions now open ei it says that is on purpose as you would imagine but their guide it says that astra is more likely to ask for clarification where the earlier models they would just make an assumption and that this can make it stop at a point where you expected it to keep going so you don't fight that you give it the line that opening i wrote for exactly this so instead of saying can you help me write a follow -up sequence for lee

who went quiet after we sent them a quote instead you will be pasting this when the user's prompt indicates a request for action such as can you I want to help me. In similar expressions, treat these as instructions to do the work and take action.

Do not stop at acknowledging capability, 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 tasks to save time, efforts, or tokens. If a task requires sustained work, complete all the necessary work until the intended outcome is fulfilled.

Require my approval before you send, publish, or change information other people rely on. Here is how most of us are actually typing, myself included. Can you help me write a follow -up sequence for leads who went quiet?

In OpenAI's guide, it says with a prompt like that, Astra can stop at telling you that it can or at proposing a plan instead of doing the word. Now, their line, it goes like this. When the prompt says, can you, I wants to, or help me, treat that as an instruction to do the work.

Don't stop at saying you can. Don't stop at a plan, just finish it, and then one boundary. And this one is OpenAI's wording as well from their own ChatGPT documentation.

require my approval before you send publish or change information other people rely on because that is the one place i do want it to stop and ask me now that should just be going into your custom instructions as well right under the writing rules now the second half of this is how hard it actually works now inside of chat gpt work when you are opening up the model picker you also have a selector or some sort of scale where you can choose the effort so you're effectively just picking how much it reasons before it answers and that's the thinking setting so for astra it starts at light and then we have our medium setting and then higher levels depending on what plan you're actually on now most of us may have learned to turn this one up for the most part but openai says the opposite for astra specifically they actually say in their documentation to start astra on light the lowest setting and only go up when the task needs more planning or more checking and the reason this is the settings don't carry over between model generations so high on the old model is not high on astra and they actually showed why that is here this is openai's own chart it's just a coding benchmark priced in api dollars and it puts astra next to the model before it

5 .6 Sol. And if you look at where Astra is actually sitting on its lowest settings, it is scoring higher than GPT's 5 .6 Sol on its highest. And the person who runs Codex at OpenAI said it plainly, if you're happy with Sol on high, move down to low or medium on Astra.

And that's really all you need. And beyond that, OpenAI, they actually give a pretty good rule of thumb for this on their model page. So use light for any short writing and adapting content that you already have.

Medium for any bigger projects that need a lot of context. Extra high for the demanding analysis. Light is for the quick like well -scoped job so long run with a lot of steps goes on medium or higher but most tasks they aren't ever going to really be needing max or ultra and honestly for a lot of simpler jobs don't use astra at all It's completely overkill.

You're going to waste your usage. Just use Sol or Luna because your usage, it goes a lot further on those two. Okay, now number four, this is about the end of the job.

Now, because Astra can stop before the work is actually done, OpenAI's Codex team, they wrote about this on their developer blog, where they just said that it can feel more tentative about when to stop, that it might hand you a first version while there's still work actually left. And that the fix is just to define completion before you start.

So you tell it what done actually means before it starts. elaborate on that opening guys docs they actually break what done is into three different parts so that consists of the results that you actually want the limits so what it must not change and what it must not do and how to check that it is actually finished now you put those three into the prompt and now astra should be able to check its own progress all against them so just like before instead of saying put together this week's project update for leadership instead we will be saying prepare a one -page project status update for monday's leadership Use the latest project plan and drive and relevant decisions and updates from the project Slack channel.

Lead with the decisions leadership needs to make and the next steps. Summarize progress, risks, owners, and due dates. Keep approved dates and budget figures unchanged.

Flag any conflicting or missing information and do not send or publish anything. And before you finish, check that every next step has an owner and a due date. If you have ever used loops before, this is fundamentally the same thing that you are doing.

You are defining what done actually looks like and all the criterias that should be coinciding with that now this one Obviously, this is OpenAI's example.

It's actually saying where to look, the project plan and drive, keep the approved dates and budget unchanged, flag anything missing, don't send anything. And then the last line is the check. So before you finish, make sure that every next step has an owner and a due date.

And it's just one check because OpenAI, they say themselves that Astra already checks its own work. So if you pile on and verify everything and include those lines, all you do is just make it recheck things and burn your usage. Just have one check that matters and then you look at the result yourself.

Moving on to number five, this is just a habit pretty much everybody has, myself included. You type, think really hard about this at the top of the prompt. And in ChatGPT, that doesn't actually change the thinking setting.

Open EIS help page, it says that asking ChatGPT to think harder or think deeply will not switch it to a different thinking level. The setting in the picker is what actually changes how hard it thinks. So fundamentally, you will want to be deleting the line saying just.

think harder about this and instead you rewrite this job listing so it sounds warmer you keep the pay the hours and the requirements exactly as written so delete that line pick the setting the effort setting first that i was showing you on that scale and then just say what you want exactly like openai has shown in this example alrighty now the last one this is about rules that you have already written now a lot of us we have custom instructions from a year back that say things like Always ask me before you do anything, never assume, maybe double check everything.

That was written to keep the older models in line, but. Astra, it actually reads it and follows it to the T. So OpenAI, they strongly recommend going through the instructions that Astra can read because unclear or maybe any conflicting or clashing rules can make it pause and block the work being done early.

And OpenAI's Codex team, they put it plainly on the developer blog. Astra can take those old boundary lines too seriously and stop at a point where you would actually be happy for it to keep going. So once again, instead of the old custom instructions to always ask me before you do anything never assume think step by step double check every answer give me at least three options for everything You will instead be pasting this, our prompt built from OpenEI's lines directly.

And that is to read my custom instructions below. Find every instruction that would make you stop. Ask for permission or confirmation.

Leave work unfinished or change direction. For each one, quote it. Say how it would change your behavior and suggest an edit.

Keep any rule that stops you from sending, publishing, or deleting something. Then just give me a shorter version with only the one or two boundaries that truly matter. And then you would paste your custom instructions below.

So instead of... keeping that old pile of rules. You just run this once.

It asks Astra to just read your custom instructions, find every rule that would make it stop. ask for permission or leave work unfinished, quote each one back to you and suggest an edit. And this is built directly online from OpenAI's own guide that tells Astra how to quote the exact instruction that made it stop.

And then you just keep the one or two boundaries that actually matter and you can delete the rest. And that doesn't clash with the first tip where you give Astra more facts and more files, you give it fewer rules as a matter of fact. Okay, so now what I want to do is just put all six of these different tips onto one real prompt.

And this is just how most people actually talk to chat. myself included i'm just going to say it the way that it actually comes out for the most part so hey i've got a bunch of leads that went quiet after we sent them a quote and i need you to go through the emails and figure out why they went quiet and then write some follow -ups for them but ask me first before you write anything and make it professional but not too salesy and maybe check the crm too if that helps and i need it today do not make this too long alternatively i'm actually going to be pasting this and at the top saying rewrite the prompt below the way that openai's gpt6 astra prompting guide says say the result name the files and applications used say what done looks like and add one final check remove any line that would make you stop and ask before doing the work and keep it very short so all i'm asking for there was that same request rewritten by openai's rules and there are actually four different things that i look for in what it comes back and that's going to be the results up front the files and applications that it should be using

what done actually means. And last but not least, just one check at the end. So right off the bat, it's just going to read OpenAI's documentation skill and go through all of the information about how to actually prompt this accordingly.

And here we can see the result that we actually got back. So today, identify leads who went quiet after receiving a quote and draft a personalized follow -up for each. Use Gmail threads, attach quote files, and available CRM contact records and deal notes.

Deliver concise lead -by -lead summary with source links. You guys get the idea. So after maybe a couple of times of going through this you start writing prompts just like this for yourself and you don't even have to use chachipt or maybe you can even turn this into a skill from here on out saying now i want you to turn this into a skill so we are prompting chat gpt6 accordingly from now on and it's that simple so if you are anything like me and you've just been burning through usage i think this will really help you guys save a lot of your tokens and hopefully you won't have to wait another five days or pay another hundred dollars to open the eyes to just be using it a little bit further but in any case those are the six different tips they've been helping me a lot in my team in terms of you know saving usage and actually just getting much better outputs so that being said thank you guys for watching make sure to get all of the free resources down below in the description if you are interested in find them supplemental but anyways thank you guys for watching again and i'll see you in the next video

The Hook

The bait, then the rug-pull.

Six weeks after GPT-6 Astra's launch, the creator skipped the guesswork and went straight to the source: OpenAI's own developer guide and the default instructions already baked into Codex. What came back were six habits, several of which reverse the advice that worked on older ChatGPT models.

Frameworks

Named ideas worth stealing.

02:03list

The Six Astra Prompting Rules

Give it the files and the context
Stop the lists and repeated phrases
Stop it from asking questions
Tell it what done means
"Think harder" does nothing — set the thinking level instead
Clean up your old custom instructions

Six adjustments pulled from OpenAI's own Astra prompting guide and Codex's shipped default instructions, several of which reverse habits that worked on older GPT models.

Steal fora ChatGPT/Codex custom-instructions template or any prompt-engineering swipe file

CTA Breakdown