Sora 2 Prompting Guide

0:00Today on the AI Daily Brief, five ideas for Sora businesses and how to prompt the most out of the Sora 2 model. Before that, in the headlines, 100 % of coders at NVIDIA are now using Cursor or another AI coding platform. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:23All right, friends, quick announcements before we dive in. First of all, thank you to today's sponsors, Super Intelligent, Robots & Pencils, Notion, and Blitzy. To get an ad-free version of the show, go to patreon.com slash AI Daily Brief. You can also now get it with Apple podcast subscriptions. However, I did talk to a contact at Spotify who suggested at the moment for how we're using it, it makes more sense to focus on the connection between Spotify and Patreon. So if you are a Spotify consumer, it looks like there will not be a native subscription version in Spotify just yet. One good piece of news is that I did learn how to connect video to the ad-free feed in Spotify, so you'll be able to get the video, not just the audio versions, ad-free.

1:03Lastly today, a quick reminder about the AIDB extension survey. As I mentioned, no changes are coming to this main show. I'm just thinking about a few possible extensions for next year, an enterprise edition, an operator's cut that's more focused on practical advice, and an investor's edition. And I'm asking people which of these, if any, are most interesting. You can take the survey at aidailybrief.ai, and I would so appreciate it if you do. But now, let's dive into the show. Welcome back to the AIDB headlines edition, all the daily AI news you need in around five minutes. I remember when, surprisingly, engineering departments and coding groups were some of the most shockingly resistant to AI inside the enterprise.

1:38Now, a year later, oh my, have things changed? Let's talk first about an update to Cursor, and then we'll talk about how this is all playing out. Cursor has introduced Plan Mode, allowing their coding agents to create plans, conduct research, and work for longer. Plan Mode gives the model new tools to create and update plans, as well as an interactive editor to modify plans in line. The way that it works is basically when you prompt the agent to create a plan, Cursor then goes and researches the code base to find anything that it needs to execute that plan. Looks through relevant files, finds review docs, asks clarifying questions, and then ultimately when you are happy with the plan, it memorializes it in a markdown file that you can then edit directly.

2:16The company said that plan mode had significantly improved code generation in their testing. AI educator Kevin Kern posted, the new plan mode in Cursor is well thought out. It builds on the existing task lists, showing they iterate on what's already there and bring features together instead of shipping separate ones. I like that approach. Investor Charles Williamson writes, the more I use plan mode and cursor, the more I realize how much attention to detail went into it. So good. Now, generally, this new feature showcases a huge emerging theme that context is a king when it comes to advanced AI use.

2:46Planning basically allows the agent to work with the user to build and refine that context before it gets to work. Now, another growing theme around AI coding is that these tools are much more capable in production now. In the first half of the year, it was surprising to many to hear that tech giants were already using AI to generate up to 30 % of their code. Yet it wasn't all that surprising to anyone that at OpenAI Dev Day this week, Sam Altman said that almost all new code written at OpenAI today is written by Codex users. Still, even with all that, it was big when Jensen Huang, the CEO of NVIDIA, the biggest company in the world, dropped this little nugget on CNBC.

3:20My favorite enterprise AI service is Cursor. Cursor is an AI coder. And every one of our engineers, 100%, is now assisted by AI coders. And our productivity has gone up incredibly. 100%. Perhaps this is why there is increasingly a push to drop the Vibe from Vibe Coding. Turns out this is just coding now. Speaking of Cursor, AnySphere, the creator of Cursor, is fielding offers of new investment at a$30 billion valuation. This rounded more than triple the valuation from the last round, which closed in June. The information reported that some potential investors have already bought shares in the secondary market at that valuation.

4:04Now, the thing that's most interesting about this to me is not that it's one more piece of evidence about how hot deals remain in the private and public markets for AI companies, but the growing conversation about context and about all of the unique data that Cursor and companies like it have that isn't available to the foundation model companies. I think people are increasingly appreciating that that ground level usage data is going to be extremely valuable when it comes to the next generation of models. And I just think that that's showing up in that lofty valuation. Another big area for rapid development in AI is around computer use.

4:37Basically AI that can use computers in the same way that humans do. Google this week previewed their new Gemini 2.5 computer use model. The model is optimized for navigating human-focused interfaces like web browsers and powers agentic features in AI mode and Project Mariner. Google said that the model outperforms leading alternatives on multiple web and mobile benchmarks. However, unlike computer use modes from OpenAI and Anthropic, Google's version is currently locked to web browsing only. The model is now available through the API, so will be accessible for developers, and BrowserBase has also spun up a demo environment.

5:09Their testing confirmed state-of-the-art capabilities, but they cautioned, computer use is hard to evaluate. You need reliable browser infrastructure and realistic tasks. Separately, Google has released a code security agent called Codemender. The agent is designed to automatically create and apply security patches. Google has been testing this agent in the open-source community for six months, having it monitor codebases and security alerts. They said that the agent has, quote, already upstreamed 72 security fixes to open-source projects, including some as large as 4.5 million lines of code. Moving over to the bidness world, another big partnership deal for Anthropic.

5:43That company has partnered with IBM to make clawed models available through IBM products. The first integration will be IBM's IDE with plans to add more soon. As part of the deal, IBM will also provide corporate education products designed to help businesses build and use AI agents based on Anthropics technology. The education will include guides on how to utilize MCP servers in the enterprise. None of the deal terms were disclosed, but Mike Krieger, Anthropics chief product officer, said, IBM knows how to navigate enterprise barriers. They understand existing tech stacks, have deep consulting capabilities, and can help with change management at scale.

6:15This partnership combines our AI capabilities with IBM's enterprise expertise to make adoption happen where it matters most. This is the second big enterprise win for Anthropic this week, after announcing on Monday their largest ever deployment, bringing Claude access to 470 ,000 workers at Deloitte. Lastly today, XAI has released a big new update to their video generation model, upgrading Grok Imagine from version 0.1 to 0.9. The Imagine platform was launched over the summer, but is getting a refresh with the much more powerful image-to-video model. The output is now much higher quality, capable of accurately producing high-motion videos and emulating tricky features like reflections.

6:50It also adds native audio and dialogue generation paired to the video clips, which of course was the huge unlock when VO3 first came out a few months ago. People are liking the results so far, the obvious comparison being how well it performs relative to Sora 2. In some instances I've seen, they're fairly close, but in others, there's still a fairly big difference. Still, XAI have made the app completely free and plugged it into X as a way to share on social media. In fact, you can use Grok to instantly animate any image on X. If nothing else, it is very clear that we are absolutely headed into the era of inexpensive, quick, and easy-to-generate video.

7:24That is, of course, why some people are very excited about all the rumors surging that we are about to get VO 3.1. It seems to be showing up in tools like Higgsfield, And so now the only question is whether coming soon means this week or in the next couple of weeks. However, speaking of AI video, that is the subject of our main episode to which we now turn. Today's episode is brought to you by my company, Superintelligent. You've got a hundred what-if ideas, but which one becomes an agent? Superintelligent maps every AI use case across your company and helps you create an agent plan that you can actually execute.

7:57We match opportunities to your tech stack, your data profile, and your team. No more guesswork, just a clear path from pilot to production. If you want agents that deliver business outcomes, start with planning. Go to bsuper.ai and sign up for a demo. AI changes fast. You need a partner built for the long game. Robots and pencils work side by side with organizations to turn AI ambition into real human impact. As an AWS-certified partner, they modernize infrastructure, design cloud-native systems, and apply AI to create business value. And their partnerships don't end at launch. As AI changes, Robots and Pencils stays by your side, so you keep pace.

8:35The difference is close partnership that builds value and compounds over time. Plus, with delivery centers across the U.S., Canada, Europe, and Latin America, clients get local expertise and global scale. For AI that delivers progress, not promises, visit robotsandpencils.com slash AI Daily Brief. Chatbots are great, but they can only take you so far. I've recently been testing Notion's new AI agents, and they are a very different type of experience. These are agents that actually complete entire workflows for you in your style, and best of all, they work in a channel that you already know and love because they are purpose-built Notion super users.

9:12Notion's new AI agents completely expands the range of what Notion can do. It can now build documents from your entire company's knowledge base, organize scattered information into organized reports, basically do tasks that used to take days, and get them complete in minutes. These agents don't just help with work, they finish it. Getting started with building on Notion is easier than ever. Notion agents are now your very own super user to help you onboard in minutes. Your AI teammates are ready to work. Try Notion AI for free at the link in our show notes. This episode is brought to you by Blitzy, the enterprise autonomous software development platform with infinite code context.

9:46Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Enterprise engineering leaders start every development sprint with the Blitzy platform bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles code for each task. Blitzy delivers 80 % plus of the development work autonomously while providing a guide for the final 20 % of human development work required to complete the sprint. Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org.

10:22Blitzy is providing a limited-time, 30-day free proof-of-concept for qualifying enterprises. The team will provide a 5x velocity increase on a real development project in your org. Visit blitzy.com and press book demo to learn how Blitzy transforms your SDLC from AI-assisted to AI-native. That's blitzy.com. Welcome back to the AI Daily Brief. One of the challenges with this show is to figure out the balance of technical news, i.e. model updates, new features, new platforms, new ways that people are using these tools, with big picture societal updates, which could include markets or geopolitics, with the more practical stuff, which is how to use and actually get the most out of AI.

11:02This, by the way, is why I have that survey going on the website, aidailybrief.ai, about what sort of extensions you're most interested in. But regardless, we have definitely been in the big picture macro realm for a little bit. And so I wanted to bring it back to the practical for today's episode with a Sora 2 prompting guide that we're expanding out to also include some examples of how people are using Sora in business. This was inspired by OpenAI dropping their Sora 2 prompting guide on their cookbook website, cookbook.openai a little earlier this week. I'll include a link to this in the show notes.

11:31And what I've tried to do is simplify and consolidate a bunch of the lessons in there to make it more easily digestible for you. Now, to just do the slightest little bit of background, Sora continues to be a very, very popular new app. Bill Peoples from OpenAI tweeted on Wednesday that Sora hit 1 million app downloads in less than five days, which he said was even faster than ChatGPT did. He said there were more features and fixes to over-moderation on the way. VC Amy Wu Martin writes, Sora is approaching 500 ,000 daily active users and holding number one in the iOS app store. The question on everyone's mind is whether retention will hold.

12:04She basically thinks that it's going to come down to how much the remix type features really differentiate it and make it feel fundamentally different than the existing short form video apps. Basically, do they give it enough lift to break out and be its own thing? Meanwhile, over on the societal side, we are starting to adapt to a world with ever more high-quality video. A bunch of Sora 2 watermark removers have started to proliferate, making that one little layer of protection that OpenAI gave people null and void for people who are willing to take the time to remove that watermark. Writes Professor Ethan Malek, I think people are still unprepared for a world where you cannot trust any video content despite years of warning.

12:42Even when Google and OpenAI include watermarks, those can be easily removed, and OpenWeight's AI video models without guardrails are coming. There are also the questions of copyright that remain. In recent interviews, Sam Altman has said that IP holders are showing up in droves to get their IP into Sora. In fact, he says that they have people who are complaining that their IP isn't being used enough. Although, of course, the flip side is also true, where there are tons of people who are very upset about their IP being opted in by default, feeling like that's a break with copyright norms of the past.

13:10For as many people are using Sora, it's kind of getting rating bombed in the App Store. I think it's down around a three right now, which who knows exactly what signal that's trying to send. Is that frustration with the app itself, or is that a general protest against AI and AI video? From a practical standpoint, Sora 2, the model, is showing up in places other than just the Sora app as well. You can access it through Pika, Higgsfield, InVideo, meaning that more and more of those videos are proliferating. And what's more, since the Dev Day announcement on Monday, Sora 2 and Sora 2 Pro have been available in the API, with people starting to build tools that make it easy to storyboard and build out more full-length videos.

13:47So before we get into the prompting guide, let's talk about a few different ideas for how to use Sora for business. While a lot of you are personally passionate and tinkers on your own terms, I know the majority of you are thinking about AI from a business perspective in some way or another, and there are a few business ideas that stand out to me as highly opportune both with Sora and on Sora the app. Within Sora, one that came straight from DevDay was using it for product design. OpenAI showed how Mattel, the toy company, had been using it to prototype new toys. After one of their designers created this sort of visual concept art of a cool road track type setting for small cars, Sora was able to turn that into an actual video, giving them a sense of just how cool it could be and seeing it in more higher resolution without having to commit the resources to actually build and prototype the thing.

14:33This seems like an amazing use case and one that I expect we'll see lots more of. And while at first I thought this might be more relevant just for big companies who are doing things like designing tracks for Matchbox cars. Who knows, man? There are a ton of independent Etsy creators out there who do everything from jewelry to pet accessories to you name it. So this is a use case that I'm super bullish on and think is going to be really interesting. A more general category is just using Sora to sell items. Basically to place products on people or in a setting in a way that brings them to life.

15:03Here's an example that Nir posted on Twitter. My boyfriend literally doesn't know how to dress. So we're using this app to help him try on different outfits. Oh my God. Should I buy this Balanchega? Nir writes, Paying a dollar a video and being able to endlessly customize the people and messages to hit different niches is pretty insane. And I think what Nir is getting at that's important is that this is not just about being able to create a good video. It's about being able to create a ton of videos. I.e., the Doctor Strange way of working comes to video production as well. Mike Fuchsia talked about a system that connected the Sora 2 API and automated workflow generator N8N.

15:40He writes, this AI system creates unlimited UGC videos using N8N in the new Sora 2 API. Fully automated, zero watermarks, HD quality. Game changer for e-commerce brands and creative agencies scaling content production. Most teams spend$10 ,000 a month on influencer content, but now with Sora 2 API, drop a single product photo, generate 50 plus HD videos with zero watermarks, own full commercial rights and pay a few bucks for video. He gave an example of a video for Groon's Gummies. Jacob Klug said something similar, although in his case, he was broadening it out. He writes, Sora 2 plus Lovable is the new startup stack.

16:15The old launch strategy, hire a 10K a month video agency, wait six weeks for launch video, pay 5K for more social cuts, get three videos that look like everyone else's. The Sora 2 reality, generate 100 launch video variations, test all the angles in 48 hours, find what converts, and ship organic content daily. Once again, you see the Doctor Strange approach of being able to just generate far, far more content and actually test it in the real world to see what's going to work. By the way, for those of you who don't know what the heck I'm talking about, if you do a search either on Google or just in my podcast feed for Dr.

16:45Strange Theory of Agent Work, it's all about how in the future, I think we're not gonna see a one-to-one replacement for the work that gets done now. I think that agents are gonna do 100 or 1 ,000 or a million times the things that we do now in a repeat process, test all of those examples against synthetic or real audiences and help us see what performs with real evidence rather than just guessing. So those are business ideas with Sora, using Sora, but not in the Sora app. But of course, to the extent that the Sora app sticks around, it is a new platform or medium for people to take advantage of and build new audiences on.

17:15One of the interesting gambits for content creators is how much to invest in a new platform. The upside is that if that platform sticks, at the very beginning, they can come in and generate an audience before other people are there, which gets them into the recommendation cycle and gives them a leg up and potentially gives them a real opportunity. in the future, again, should that network persist. Now, the downside is if it goes away, all that effort was for nothing. I think of the folks that I knew who grew hundreds of thousands or even millions of followers on Clubhouse, only for that entire network to go the way of the dodo.

17:44Still, I think there's a couple ideas that seem pretty obvious and interesting to me on the Sora app itself that I think are going to be interesting for people to experiment with. The first, unsurprisingly, is education and tutorials. I've been messing around with this a little bit, although not with super strong intention. Hey, future thinker! Ever wonder what makes AI Answers smart and not just lucky? It's all about context engineering. Context is the who, what, where, and when that surrounds your data. Reinforcement learning is like teaching a game player. It tries moves, sees what scores, and gets better through rewards.

18:16Think of it as trial and error with high fives from the future. Another related category is more little news briefs. This one comes from Draquez, who, among a million other businesses, works a little bit with us here at AIDB.

18:38And then the one that I think is just the most default that people are going to have a lot of fun with when they get there is making unhinged commercials for your small business. So this one actually isn't for a particular business. It's from Justine Moore, A16Z. But imagine that this was a video for a local bar and they published it both on Sora, but also to their other networks as well.

19:06Despite the fact that the bartender is a giant capybara, I very much want to go to this establishment. All right, so how do you get the most out of this new platform? What are the prompting tips that are going to drive impact for you? As I mentioned, a lot of this is coming direct from the horse's mouth, by which, of course, I mean the Sora 2 prompting guide, which OpenAI dropped earlier this week. There is a ton of detail in here. I'm not going to try to get to everything. I'm going to try to hit the big picture points in a way that's more accessible. And the first one, I think, is when to know to be open versus descriptive when it comes to your prompts.

19:39What I mean by this is that it's our natural instinct, I think, to try to provide the AI a ton of really specific detail to get exactly what we're seeing in our mind's eye out of it. And frankly, figuring out how to do that is going to be a lot of what these tips are about. However, there is also another type of use where you just let the AI be more creative. Sometimes it's going to be valuable, especially if you're not exactly sure what you're trying to get outside of a few key details, to kind of just let the AI do its thing. As OpenAI writes, shorter prompts give the model more creative freedom, and you can expect more surprising results.

20:11Now, short prompts can be super short. For example, I just said a fun 1980s educational video explaining reinforcement learning in two sentences. Or they can be a little bit more detailed. The example of a short prompt that OpenAI gives is, in a 90s documentary-style interview, an old Swedish man sits in a study and says, I still remember when I was young. The reason they say this will work well is that 90s documentary sets the style of the video, old Swedish man sits in a study, describes the subject and setting, but in a minor detail, which, as they say, lets the model take creative liberties, and by giving the dialogue, I still remember when I was young, Sorrel will be able to follow that exactly.

20:48They say that this prompt will reliably produce videos that meet these requirements, but all the stuff that's left out, i.e. the prompt not describing the time of day, the weather, the outfits, the tone, the look, the age of the character, the camera angles, the cuts, the set design, because the prompt doesn't get into those things, they're left to the AI. And the point is that sometimes leaving those things up to the AI is going to produce really wonderful and valuable results that are different than if you had constrained every little thing. Now, let's expand this out a little bit and try to move into a realm that's slightly more detailed.

21:19One thing that OpenAI suggests is to think about each shot as its own unit. Basically, they say that you can put multiple shots into a single prompt, especially if you have a sequence of action that needs to happen. But if you do this, think about each shot as a unit. And what does that mean? What's in a unit? A strong unit in a prompt is going to have one of each of the following. First, a style reference. So for example, that one we heard, 90s documentary, from the piece that I showed you before, 1980s educational video. That is the style reference. And I think in a lot of ways, just for getting the model in the ballpark of what you want, this strikes me as one of the most important pieces.

21:57So much of video is about the vibe that you're trying to set. And the style reference is a really high impact and shorthand way of helping the AI achieve what you're looking for. Next up, you've got the camera setup, which could be what type of camera it is or where it's framed relative to the subject of the video. One subject action. They recommend that in the context of this one unit, whatever your subject is, only has one main action they're taking or one main thing they're doing so as not to overwhelm the unit. That doesn't mean that a single prompt can't have multiple subject actions, but again, they're trying to bundle these things in units that then get added together.