GPT-6 Astra for Meta Ads: 8 Things Your Ad Tool Could Never See

8 things your ad tool could never see — and how to run each one

You commented "astra." Here's the whole thing.

Every tool you've used for ad research reads text. Copy, CTA, landing URL, spend, days running. Useful, but it's the data around the ad. Nothing that makes an ad work is in there — the first frame, the face, the text burned on screen, whether it looks shot on a phone, whether the landing page matches.

Astra can look. It scores 92.7% on ScreenSpot-Pro, up from 76.9% for the previous model, and it reads a screen the way you do. This guide is eight jobs that were impossible before that, each with the exact prompt, what good output looks like, and — because seeing isn't making — what to do with the finding once you have it.

That last part matters. Every job here ends with a brief. A brief is only useful if something turns it into a creative. I'll be straight about where that happens: it's the reason we built HeyOz , and you'll see it at each handoff.

What Astra computer use actually is

It's a plugin in the ChatGPT desktop app that lets the model see and operate your screen. It runs on your machine, in your browser, with your signed-in sessions. That's different from cloud agents that spin up their own VM — and it's why this works for ad research without handing credentials to anyone.

Install: ChatGPT desktop app → Plugins → Computer Use → Install. On macOS, grant Screen Recording and Accessibility when prompted. On Windows it runs in the foreground — it takes over your pointer and keyboard while a task runs, so plan to step away or use a second machine.

Invoke it by writing

or in your prompt, or just asking ChatGPT to use Computer Use. Select GPT-6 Astra in the model picker for anything that depends on visual judgment — which is all eight of these.

Approvals: the first time it wants to use an app, it asks. You can choose Always Allow per app. Do that for Chrome. Don't do it for anything else.

If the plugin isn't showing, check your plan and region settings — availability varies.

Three rules

1. Ad Library in the browser, your ad account through the API. The Ad Library is a public page, so browsing it carries no account risk. Your own Ads Manager is different — automating clicks inside it risks restriction, and Meta has a full Marketing API for anything you'd want to change. Every prompt here that touches your account is read-only and visual.

2. Never let it post. Instagram, Facebook, LinkedIn, TikTok and X all restrict automated account access. Research from public pages is fine. Publishing from your logged-in accounts is not.

3. Treat pages as untrusted. OpenAI's own docs say it plainly: if ChatGPT uses your browser it can act on pages where you're signed in, and pages can contain misleading content. Keep sensitive tabs closed, keep tasks narrow, and stay present for anything that touches account or payment settings.

The context file

Create a note — a text file, a doc, anywhere — and paste it at the top of every session. Astra's output is generic without it and sharp with it.

That last line is what makes the whole system work. More on it in Part 5.

Four jobs that read what the category is doing visually. All four use the public Ad Library.

Job 1 — First-frame teardown

What it does: Freezes every competitor video ad at second zero and logs what's actually in the frame.

Why: The first frame decides whether the rest of the ad exists. Everyone knows this. Nobody has systematic data on what their category's winning first frames contain, because it's a look-at-it question.

Prompt:

Good output looks like: 20–40 rows, three or four percentages, and one sentence that tells you something you didn't know — "every ad over 45 days opens on a hand holding the product, no faces."

The handoff: You now have the first-frame formula. It's a paragraph. Turning it into an actual first frame — in your brand, with your product — is the job HeyOz does: paste the brief, and it renders the shot in your typeface and colours rather than an approximation. Run three variants on the formula and test them, the same afternoon.

Job 2 — Visual hook archetypes

What it does: Clusters ads by what they look like, not what they say.

Why: Two ads with completely different copy often run the identical visual play — a hand entering frame, a split screen, a text-over-black open. Copy-based tools can't see that. This can.

Prompt:

Good output looks like: six to ten named archetypes, a ranking that surprises you at least once, and a gap.

The handoff: The gap is the brief. Take the unused archetype, write one line describing the visual move, and hand it to HeyOz with your product URL. That's an ad nobody in your category is running, in your brand, ready to test.

Job 3 — Text density scoring

What it does: Measures on-screen text across every competitor ad and lines it up against survival.

Why: "How much text should be on screen" is a permanent argument that nobody settles with data. This settles it for your category.

Prompt:

Good output looks like: a clear band — "long-runners sit at 4–9 words per frame, lower-third, caption-style" — or an honest "no pattern," which is also useful.

The handoff: Density and position are two lines in a brief. Every format on HeyOz takes them as inputs, so the number you just found becomes the number your next creative is built to.

Job 4 — Brand system extraction

What it does: Pulls a competitor's entire visual system — colours, type, layout logic, pacing — across their full active set.

Why: You've seen their ads one at a time. Seeing the system tells you what they've standardised, what they've stopped doing, and where you can look nothing like them.

Prompt:

Good output looks like: a brand doc for a company you don't work at, plus three differentiation moves.

The handoff: Here's the thing worth noticing. You just made Astra extract a brand system by looking at ads for twenty minutes. HeyOz does the same thing to your brand automatically — from your URL — so every creative it produces is in your system, not a near-miss of it. Point the three differentiation moves at it and you get ads that look like you and nothing like them.

Three jobs that look at your account the way strangers do. All read-only.

Job 5 — Your ad as it actually renders

What it does: Shows you your ad in the feed, on a phone, with everything Ads Manager's preview hides.

Why: The preview is a lie of omission. Real placement truncates the caption, attaches your profile picture, and puts your ad directly between the two ads you're losing to. That's what the customer sees.

Prompt:

Good output looks like: a short, uncomfortable list. "Your hook line gets cut at word six. Competitor A's survives. Your product is illegible at thumbnail size."

The handoff: Use this as your QA step. Whatever you generate on HeyOz , run it through this prompt before spending on it. Fix the truncation in the caption, regenerate the first frame if the product's lost at thumbnail size, and only then launch.

Job 6 — Safe zone audit

What it does: Checks every Reel you're running for text hiding under the interface.

Why: Reels overlay the caption, the CTA, the audio badge and the right-hand icon stack on top of your video. Text placed under any of them is invisible. It's free to fix and almost never checked.

Prompt:

Good output looks like: either "clean" or a short list with timestamps. Both are worth knowing.

The handoff: When you regenerate, brief the safe zones explicitly — "all text in the centre 65% of frame, nothing below the lower fifth." HeyOz will build to that constraint, and the audit becomes a one-pass check instead of a recurring fix.

Job 7 — UGC vs produced

What it does: Settles whether phone-shot or produced creative is winning in your category, with data.

Why: Everyone has an opinion. The answer changes by category and by year. This is the only way to know for yours.

Prompt:

Good output looks like: a number and a confidence level. "3s and 4s dominate the long-runners; pure UGC dies fast here; moderate confidence, sample of 31."

The handoff: Both ends of the scale — talking-head UGC and produced product film — are formats HeyOz generates from the same product URL. So when the data says "3.5," you can produce exactly that register and test it against the ends, rather than picking one style and hoping.

Job 8 — Ad-to-landing-page coherence

What it does: Screenshots the ad, screenshots where it lands, and judges whether they look like the same brand.

Why: This is one of the largest conversion leaks in paid social and it's almost never audited, because it's purely visual. A person clicks a warm, phone-shot ad and lands on a cold corporate page. They don't consciously notice. They just leave.

Prompt:

Good output looks like: a weakest pair and a clear diagnosis. Half the time the fix is the ad, not the page.

The handoff: When the fix is the ad, it's usually that the creative drifted from the brand while the page didn't. Regenerating on HeyOz pulls the ad back to your actual brand system — the same typeface and colours the page is already using — which closes the gap from the ad side without touching the site.

Here's the honest structure of everything above.

Each of the eight jobs ends with a brief. A first-frame formula. An unused archetype. A density range. A register on the UGC scale. A mismatch to fix. Astra is excellent at producing these, and it produces them in minutes.

None of the eight produce an ad.

That's not a limitation to work around — it's just where the tool stops. Computer use is for reading and judging screens. Producing a brand-accurate creative from a brief is a different problem: it needs generative models, brand extraction, and a rendering pipeline. Astra doesn't have those and isn't meant to.

So the loop is:

The research half runs at machine speed. If the creative half runs at human speed, the whole loop runs at human speed and the research spends most of its week waiting.

HeyOz is the creative half. Product URL in. It pulls your brand — typeface, colours, spacing, product imagery. You give it the brief Astra just wrote. It renders the ad: video, static, UGC-style, carousel. Same day. In your actual brand, not a model's guess at it.

The handoff template

Every job's closing paragraph should slot into this. Paste it into HeyOz as the brief.

That's a complete, evidence-backed creative brief. It took Astra about an hour of screen time to fill in, and it would have taken you a full day of scrolling and squinting — and you'd have been guessing at half of it.

Don't run all eight every day.

Monday — Jobs 1, 2 and 3 on the competitor set. Thirty to forty minutes of screen time. Produces the week's angle.

Tuesday — Brief into HeyOz, generate three variants, run Job 5 on the outputs, launch.

Thursday — Job 8 on everything live. Fix the weakest pair.

Monthly — Job 4 on one competitor, Job 7 on the category, Job 6 on everything running.

That's roughly two hours of Astra time a week producing a tested creative every week, with the research fully documented.

It's describing instead of counting. Your context file's "counts, not adjectives" line isn't in the session. Paste it again.

Ad Library returns nothing. Search the competitor's exact Facebook Page name, not their brand name. They're often different. Check the country filter matches your geo.

It's clicking around inside Ads Manager. Stop the task. Every job here works from the public Ad Library. If you want account data, use the Marketing API or an ads connector — never browser automation on a live account.

Windows: it keeps taking my mouse. That's by design on Windows. Run it on a second machine, in a VM, or step away for the duration.

It says it can't tell. Good. That's the honest answer and it's better than a confident guess. Narrow the question or accept the uncertainty.

After a month: you know what the winning first frame looks like in your category, which visual plays are crowded and which are open, the text density that survives, the register that wins, and whether your landing pages are quietly undoing your ads. All of it from evidence rather than instinct.

And you'll have a stack of briefs.

Whether those briefs turn into ads the same day or sit in a doc for a fortnight is the only thing that decides whether this system compounds. Astra sees. HeyOz makes. Both halves, or the loop doesn't close.

Reply with what Job 1 finds in your category. I read all of them.