Last week I walked you through ChatGPT Work — the agent side of the platform and how to start putting it to use. This week I wanted to go a level deeper, into the model actually underpinning it. Understanding how Astra is different is what lets you get the most out of Work and the tasks you hand it, so it's worth the time even if you're mostly living in the Claude ecosystem right now.

Because this is a genuine step-change from previous ChatGPT models — different enough in what it can take on that the old habits don't get the best out of it — I wrote a full guide on what's new, how to use it, and how to prompt it. This is the short version; the complete thing is linked at the end.

The interesting thing about GPT-6 Astra isn't that it writes a marginally better product description or ad. It's that it can take on a larger job — gather information, work across files and tools, make decisions, produce finished assets, check its own work, and keep moving through a multi-step workflow.

For ecommerce teams, that shifts the useful question. It's no longer "what prompt gets me a product page, an email, or five ad variations?" It's "what complete workflow could I hand over — from inputs to a checked, usable result?"

The mental model that keeps this straight:

That last point matters. A more capable model doesn't remove the need for clean data, permissions, checks, and human judgment. Astra can't read your Shopify, Amazon, Meta Ads, Klaviyo, or inventory system unless the access is actually there.

A few changes are worth flagging above the rest:

It handles workflows, not just answers. A single assignment can now include research, analysis, creation, validation, and packaging. "Summarise these reviews" becomes "analyse the export, quantify recurring positives and negatives, compare them against our listing claims, identify the three highest-value messaging gaps, rewrite the affected copy, and deliver a report plus a change log."

It stays coherent through bigger jobs. It's better at holding the objective across long, multi-input tasks. One caveat worth keeping your skepticism about: this does not mean every task runs for a week. Runtime depends on the surface, permissions, usage limits, and how you've scoped the job. Treat "week-long model" claims as commentary unless your own interface explicitly supports it.

It can work through interfaces. With the right capabilities enabled, it can drive browsers and computer interfaces — not just reason about screenshots. Treat that like handing a colleague access to your screen: narrowest access needed, test environments first, and purchasing, publishing, customer comms, and budget changes stay behind approval.

It follows instructions more closely — which cuts both ways. Astra is more sensitive to your instructions, skills, and files. That's powerful, but it also means old, bloated, contradictory "master prompts" now actively make results worse. Don't paste a 2,000-word legacy prompt into every task. Keep the outcome, the real constraints, the evidence required, the output format, the approval boundaries, and the finish line — and cut the ceremonial step-by-step.

Prompting hasn't disappeared. What good prompting looks like has changed. The old approach prescribed the method — "first do X, then Y, use this exact sequence, check with me after every step." With Astra, that scaffolding is often counterproductive. A better brief defines the destination and lets the model find an efficient path to it.

This is very much process before prompts territory: the human skill moves from prompt engineering to task selection, context design, success criteria, permission boundaries, quality control, and reusable workflow design.

For any serious task, give Astra six things: a goal (the business result, not the activity), context (what you sell, where, the period, the commercial logic), inputs and evidence (the sources to use, which is authoritative, what to do when data conflicts), constraints and authority (what it may do, what needs approval, what it must never touch), output (the exact artifact you need), and a finish line and checks (what "done" means and how you'd inspect it yourself).

Here's the copy-and-use template from the full guide:

I kept this issue to the ideas that change how you think about the model. The complete guide goes considerably deeper, with:

The full guide — what's new, how to use it, and how to prompt it — is my gift to you here :)

The takeaway is simple: Astra's value isn't better prompting, it's the ability to delegate a larger unit of work. Give it a commercially meaningful goal, the right context, controlled access, and a clear definition of done — then make it produce and check a usable artifact. Start with one real recurring workflow, keep live actions behind approval, and turn what works into a reusable process.

📑 Reducto - Convert complex docs into LLM ingestible chunks.

✨ Chat Ur Data GPT - Boost productivity and unlock data insights with AI-driven assistance

🎨 Phedra - Create your own versions of any image you find online.

🤖 Sixty AI - Runs in the background of your devices, managing all your incoming messages, invites and alerts and only interrupting you when it’s really important.

😍It was great, all I needed

🙂Good, not great

🙄It sucked