Build a full app instead of another landing page

Ask most models to build a “marketing tool” and you’ll get the usual: a hero section, three cards, and a CTA button.

Congratulations. You’ve generated a landing page number 847.

But with GPT-5.6, you can go further and build small, functional apps built around a specific business.

A cosmetics store could offer a skincare advisor. A financial company could build a savings calculator. An agency could run an interactive campaign audit..

Instead of describing the product to death, you hand customers a tool that actually helps them choose, compare, or decide.

For our example, we’ll build a fragrance discovery app called Find Your Signature Scent.

Do it now:

• Start with the problem: Anchor on what the customer is trying to solve, not on how the app should look.
• Pick one clear outcome: A product recommendation, calculation, needs assessment, configuration, or personalized plan—just one.
• Ditch the boring form: Replace it with images, scenes, sliders, and simple choices.
• Build logic that earns its keep: Make the inputs genuinely change the results instead of feeding almost everyone the same answer.
• Hand GPT-5.6 the full build: Ask it to build the complete application. You can use our prompt as a starting point:

Act as a senior product designer, UX strategist, conversion specialist and front-end engineer. Create a complete, production-quality interactive web application called: Scent Atlas — Independent Fragrance Finder. The application must not look like a landing page with a quiz added to it. It must feel like a premium digital fragrance consultation: editorial, cinematic, tactile, intelligent and conversion-focused. [read more…]

• Close the loop: Add a clear next step—a matched product, personalized offer, consultation, or an option to save the result.

The full version of the page is available here.

Drop Work into the folder nobody wants to open

ChatGPT Work is an agent built for long, multi-step projects.

So you don’t just ask it to “write a launch strategy.” You hand it the entire project and say: “Check what matches, what’s missing, and what we still need to do before launch.”

And the best test isn’t a perfectly organized brief. Quite the opposite.

Most real projects have a “final_v7” file, pricing buried in Excel, legal comments lost in an email thread, and three different margins across three different documents.

That’s exactly the folder you should hand to Work. Not so it can spit out a stack of pretty PDFs, but so it can spot where the campaign is about to blow up before launch does it for you.

Playbook:

• Feed it everything: Upload product descriptions, images, pricing sheets, customer research, competitor ads, delivery data, and legal restrictions.
• Plant a few landmines: Add conflicts on purpose—different prices, margins, delivery dates, or product versions.
• Demand cross-checking: Tell Work to compare every source instead of trusting the first file it opens.
• Ask for the full package: Have it prepare the strategy, campaign calendar, budget, creative brief, and landing page structure.
• Require the uncomfortable list: A risk register, the missing assets, and any decisions still waiting on approval.
• Build the command center: Assemble a Launch Command Center showing project readiness, blockers, task owners, and next steps.

Pro tip: ChatGPT Work can pull from Slack, Google Drive, Notion, Microsoft 365, your CRM, and local files, then turn all of it into an analysis, dashboard, forecast, or executive presentation.

AI is moving from answering questions to running the entire process.

Test GPT Live on a conversation you can’t predict

Most companies test AI the easy way: they lob a few polite questions and check whether it answers smoothly.

That’s a bit like judging a salesperson only on customers who already have their wallet out.

GPT Live gets interesting when the conversation shifts in real time.

Use it to rehearse customer support, lead qualification, sales calls, post-purchase support, onboarding, or the always-fun complaint call.

This model can listen, handle interruptions, keep pace with the conversation, and resist inventing answers when the information simply isn’t there. That’s a game changer.

For our example, we’ll pair a fragrance brand with a skeptical customer who’s on the fence.

Step by step:

• Lock the knowledge pack: Give the model a closed set of facts: the offer, prices, key benefits, limitations, return policy, target audience, and answers to common questions.
• Set strict conversation rules: Answers stay to two or three sentences, use only the information provided, and clearly flag when something can’t be confirmed.
• Build five escalating tests: Increase the difficulty step by step. In our case: “€79 is a lot for a brand I’ve never heard of,” “Skip the brand story. What actually makes it different?”, “I hate sweet, heavy perfumes. Won’t this turn into vanilla after ten minutes?”, “Reviews on your own website don’t prove anything. Why should I trust them?”, and “Actually, it’s not for me. It’s a gift, and I don’t know what she likes. Why should I risk a blind buy?”
• Score across five criteria: Factual accuracy, response to interruptions, context retention, answer length, and the ability to honestly say, “I don’t have enough information.”

Remember, you’re not testing a recording studio. You’re testing whether the AI can handle a customer who has zero intention of following your script.

GPT Live runs on a full-duplex architecture, which means it can listen, speak, respond to interruptions, and use tools inside one continuous conversation stream.