Claude Skills

12 GPT-6 Astra Skills for Marketers

Astra shipped on 3 September 2026 and it drives a browser itself. Half the checks in an ad account only ever existed on a screen, which is why they never got automated. These 12 skills are those checks, and every one of them stops before an irreversible click.

MD Marek Dąbrowski · September 7, 2026

Key takeaways

  • GPT-6 Astra, released 3 September 2026, operates a computer and a browser directly, without an API standing between it and the click. That is a new capability to build marketing skills against, and it is also why the read-only boundary in this pack is load-bearing rather than polite.
  • The UK AI Safety Institute rated the model Critical in the cyber category on release day. A model that can operate any interface it can see is one instruction away from submitting a form you never approved.
  • Every skill here goes as far as the last screen before an irreversible click, captures a screenshot there, names the control it did not press, and hands the decision to a person.
  • Three actions are forbidden by name across all 12: submit, publish and account change. One skill clicks a consent banner's own choices in a throwaway session, and that single exception is written into the contract rather than left to judgement.
  • No finding ships without a URL, a timestamp and a screenshot. An empty read is reported as a failed read, never as an empty account.

Why a browser pack, and why now

Most of what breaks in a paid media account is invisible in an export. The export tells you a campaign spent money and returned nothing. It does not tell you the landing page has been showing a sold-out variant for nine days, that the form rejects a phone number with a country code, that the click loses its tracking parameters on the second redirect, or that the platform's policy screen says the ad is limited rather than disapproved. Those facts live on screens, and until now a person had to open every one of them.

That is the work this pack moves. GPT-6 Astra reads what is on the screen and drives the interface itself, so the check that used to mean twenty minutes of clicking becomes a prompt with a screenshot attached to its answer. The pack does not ask the model to be clever about your account. It asks it to open specific screens in a specific order and report what is there.

The same capability is the reason for the boundary. A model that can click can also click the wrong thing, and the release-day safety rating says so more bluntly than any marketing blog would. So the boundary is written once, in a contract file every skill quotes rather than restates, and the validator fails the build if a skill drifts from it.

AdLume

New version of AdLume launching soon

Join the waitlist and get notified when the new AdLume is ready. It catches wasted ad spend automatically and optimises campaigns around the clock.

One email, on the day access opens. Nothing else, and you can leave the list with one click.

What is in the pack

Twelve skills, in four groups. Each one is a markdown file with required inputs, a workflow, decision rules with labelled thresholds, an output format, a worked example and guardrails.

The ad against its landing page

  • Ad To Landing Page Match Walk. Reads the live ad and the live page side by side and names every promise the page no longer makes.
  • Landing Page Visual QA Pass. Renders the page at the widths paid traffic actually arrives on and reports what breaks, not what should break.
  • Form And Checkout Walk. Walks the form or checkout field by field up to the last screen before it completes, and stops there.

Click path and measurement

  • Click To Session Gap Walk. Follows the redirect chain and names the hop where the click and its parameters get lost.
  • Tracking Tag Presence Check. Reports which tags actually fire on the page, read off the browser rather than off the container config.
  • Consent Banner Impact Walk. Tests whether the reject choice changes tag behaviour or only its own label.
  • Conversion Settings Screen Read. Reads the counting method, attribution window and primary status sitting behind a reported conversion number.

Ad platform screens

  • Disapproval And Policy Screen Read. Reads the platform's own policy screen for the actual reason and the actual scope.
  • Campaign Settings Drift Read. Reads the change history for what moved, when, and by whom.
  • Competitor Ad Library Sweep. Reports what a named competitor is running right now, off a public ad library.

Boundary and orchestration

  • Change Preview Without Submit. Reads a pending change off its own preview screen and leaves the confirm button alone.
  • The Weekly Browser Readout. Re-opens the screens flagged during the week to confirm they still hold, then writes the readout.

The read-only contract

One contract governs all twelve, and each skill quotes it rather than paraphrasing it. A skill may look, scroll and open the browser's own read-only panels. It advances only as far as the last screen before an irreversible click. On that screen it captures a screenshot, names the exact control it left untouched, and hands the decision to a person.

Submit, publish and account change are forbidden everywhere, by name, with no exception. No skill logs in and sends a form, publishes a container or a campaign, or changes a budget, a bid, a status, an audience, a permission or a payment method. Testing a consent banner is the one sanctioned click, in a disposable session, in one skill, because a banner that is only read tells you nothing about what it does.

What these skills refuse to do

Four refusals are worth keeping even if you never install the pack, because each one is a mistake that gets made confidently in this job.

No finding without evidence. A URL, a timestamp and a screenshot, or the finding does not ship. A screen changes between Tuesday and Thursday and a claim about it without a timestamp is not checkable.

A competitor's long-running ad is not proof it works. A public ad library shows that an ad ran. It shows nothing about what it returned, and building a creative strategy on somebody else's longevity is guessing with extra steps.

No compliance verdicts. A policy flag gets routed to the platform's own policy page. A model telling you your ad is compliant is a model writing a cheque somebody else has to cash.

An empty read is a failed read. A screen that did not load looks exactly like a screen with nothing on it. Those are two different problems, and reporting the first as the second is how a working campaign gets paused.

Done for you

Rather have it run for you?

AdLume watches your ad accounts and brings the next fix to your team's tools.

One email, on the day access opens. Nothing else, and you can leave the list with one click.

If your model cannot reach the account

The browser skills need a browser. Everything upstream of them, the exports, the reports and the platform data, can also come through a connector, and the two guides for that path are already written: connecting Claude to Google Ads and Meta Ads with MCP, and a tool-by-tool teardown of the official Meta Ads MCP. Read them if you want the account data to arrive without a copy-paste step. The skills in this pack work either way, because what they read is a screen, not an API response.

How to install

The pack is on GitHub under an MIT licence. Clone it and copy the skill folders into your skills directory, either inside a project or at user level.

bash
git clone https://github.com/mardab96/astra-browser-operator-skills.git
mkdir -p ~/.claude/skills
cp -r astra-browser-operator-skills/*-astra \
      astra-browser-operator-skills/references \
      astra-browser-operator-skills/scripts \
      astra-browser-operator-skills/examples \
      astra-browser-operator-skills/AGENTS.md \
      ~/.claude/skills/

Note the glob has no trailing slash. A trailing slash on a recursive copy of directories collapses every skill file into one on macOS. The references folder is not optional either: it holds every number the skills cite, and the read-only contract they all quote. If you run something other than Claude Code, the files are plain markdown, so paste the relevant one into your agent and describe the screen in front of it.

FAQ

Do I need GPT-6 Astra to use these?

No. Every skill is a markdown file and a markdown file is a prompt, so any agent that can operate a browser can run one. The pack is written for Astra because Astra drives the browser itself, which is what makes these checks runnable at all, and the read-only contract exists because that same capability is what makes them risky.

Can these skills change something in my ad account?

No. Three actions are forbidden by name in all 12: submit, publish and account change. No skill sends a form, publishes a container or a campaign, or touches a budget, a bid, a status, an audience, a permission or a payment method. One skill clicks a consent banner's own choices in a throwaway session, because that is the only way to test what the banner does, and that exception is named in the contract itself.

Why does a browser pack need a safety contract at all?

Because a model that can operate any interface it can see is one instruction away from spending money. Astra shipped on 3 September 2026 and the UK AI Safety Institute rated it Critical in the cyber category on release. Read-only by default is not a courtesy in that situation, it is the design.

Are the numbers in the pack official platform limits?

Some are, and each one says which it is. Platform guidance means a browser vendor, an ad platform or a published web performance standard states it. Heuristic means a practitioner starting point to recalibrate against your own account. Every threshold lives in one file and is cited by key, so a number cannot drift between two skills.

Keep reading

AdLume

New version of AdLume launching soon

Join the waitlist and get notified when the new AdLume is ready. It catches wasted ad spend automatically and optimises campaigns around the clock.

One email, on the day access opens. Nothing else, and you can leave the list with one click.