Ship Weekly, Cut Scope, and Keep a Secret Roadmap: How Captions Builds in the AI Era

Gaurav Misra is co-founder and CEO of Captions, a consumer AI video product with more than 10 million users that has raised over $100 million. He was also an early Snap employee who led design engineering, and the host frames the conversation as an archeology of how a modern AI-oriented startup operates. Captions lets users generate and edit talking videos with AI.

Gaurav says this is a rare moment in which so much is possible: five or seven years ago it felt as if every idea had already been tried, but now things he tries just work. There are more possibilities than there are people available to pursue them, and he does not expect that window to last forever. He feels lucky to be part of the movement.

The host notes that the technology for building has arrived, but getting people to pay attention and stick around is increasingly difficult because everything is easy to build and interesting. Gaurav agrees that saying “we’ve rethought this thing with AI” can persuade people to try a product, but only if the product delivers on the promise. If users play around and leave, the hype does not become a business.

Gaurav warns that people will come for the cool factor, but a playground without a real problem is not a business. The fundamentals remain: identify a user problem, apply technology to solve it, and find a mechanism to reach people who have that problem. He also notes that science-fiction interfaces like JARVIS or TARS now exist in some form, yet adoption is so gradual that people barely react.

The new product-development problem is not a lack of possible paths but too many, so prioritization now means deciding what not to pay attention to. Gaurav says Captions looks for user demand by watching what has virality and asking what core element made it resonate. That core may be a one-time use case rather than a subscription, but it can still be identified and molded into the business; ideas can be tested in conversation and social sharing before anything is built.

The operating goal is for every engineer to ship a marketable product every week, meaning something a user might subscribe or come to the app specifically for. Table-stakes features like justify alignment in a word processor do not count, because nobody chooses a product for them; a unique capability can bring users even if obvious features are missing. The company expects much of this work to fail, doubles down on what starts to work, and treats each release as an MVP that is sliced until removing anything more would make it useless.

When time pressure hits, Gaurav says the mistake is to cut quality rather than scope. The method is to examine every element that takes time and ask what happens if it is removed, repeating until the product would become useless; that is the one-week project. His example is adding an image to a video: cut background removal, hue and saturation, cloud pickers, and custom UI until the feature is just a native camera-roll picker dropping an image into the video.

Shipping the stripped-down version gives the team a prioritized list of complaints: users will say the missing background removal is the biggest problem, and the team can ship fixes the next week. Complaints are a sign that people care, while no complaints can be a red flag. This weekly loop also keeps the company above the noise, because it constantly produces new, demonstrable features.

Captions separates long-term infrastructure work from the weekly feature cadence; Q4 is infrastructure quarter, when the team pays down technical debt as product delivery winds down. Gaurav argues that a startup’s job is to take on technical debt because that is how it moves faster than a bigger company. Large companies either pay debt immediately or spend years repaying debt taken on when they were startups; Snap, where he worked, had many examples.

His heuristic is to ask whether a problem must be solved today or whether the 50th, 100th, or 500th engineer can solve it later. If a future engineer can solve it, the startup should use that future engineer now and push the problem forward; if the company fails, that engineer will never be hired and the debt will not matter. He compares technical debt to financial debt: it creates leverage, letting a small team build products it could not otherwise afford.

The limit is interest. Every piece of debt takes some daily time in bugs, restarts, crashes, and maintenance—1% or 2% per day—and enough debt can consume 80% or 90% of the team’s time until it is only keeping the lights on. That is the failure case: a technical-debt runway runs out before the company has delivered enough value to hire engineers who can pay the interest or pay off the debt. For big decisions, he applies the two-way-door rule: move fast if it is reversible, and think carefully if it is a one-way door.

AI coding tools compound the advantage: everyone at Captions uses Cursor, the team has tried many tools, and it has used Devin as an AI engineer that can be chatted with in Slack. Gaurav says a startup can adopt such tools faster than a large company that needs many lawyers in the room first. He expects companies eventually to have hundreds of agents, including managers of agents.

Captions sometimes starts with a PM and a product spec, but often it starts with design: designers explore many ideas without knowing exactly why, then sit with PMs and look for interesting concepts that metrics-first planning would miss. The company is small across design, PM, engineering, iOS, Android, web, backend, ML, and research, and it deliberately reverses the usual process. Gaurav says this design-first path can surface unique ideas.

The public roadmap collects user requests—background removal, undo and redo, longer uploads—and prioritizes them by affected users and market. Those features are table stakes because every competitor knows about them: if a user asks Captions, they are asking everyone. The secret roadmap contains things nobody asked for and that users might initially say they do not need, but which the team believes can change behavior and make people never go back.

Secret-roadmap ideas come partly from company-wide quarterly brainstorming that includes engineering, recruiting, and everyone else, not just product. People generate and rank ideas, then the product team evaluates feasibility and technology. One early example was eye contact: new video recorders often read from an off-screen script, so Captions worked with Nvidia to shift the speaker’s eyes toward the camera; the launch video went viral worldwide and the feature was copied across the industry.

Gaurav would like LLMs to join brainstorming, but he says context is the blocker. Much of his understanding of users and use cases is abstract, and he struggles to verbalize it even when teaching PMs and designers, so making it available to a model is harder still. He points to Snap CEO Evan Spiegel as an example of unmatched user understanding: Spiegel would propose ideas everyone disagreed with, launch them, and produce hit after hit while nobody understood why.

The host notes that Snap is the last social network to launch and stick around besides TikTok, and that there has been little insight into how it operated. Gaurav says strong product-market fit can be dangerous because a social network may grow despite what the team builds, teaching people the wrong lessons about which contrarian choices worked. Snap survived intense competition partly by continuing to innovate, with a structure unlike any he has seen elsewhere.

At Snap, the design team was central and stayed small—about 10 to 12 people even at 5,000 or 6,000 employees—and for a long time there were no PMs. PMs arrived with monetization, but during the innovation period the company was designer-led; the catch is that these were not average designers, because they also did PM work such as writing documents, setting roadmaps, coordinating teams, and owning shipping schedules. The arrangement gave the CEO granular control: if you wanted to add UI to the app, it needed design approval, and design talked directly to Evan.

The designers were individual contributors with no reports, and the job was high-workload and highly paid, with compensation above what designers or PMs or engineers might normally expect. Gaurav says someone still has to do the PM work, and not every designer wants to write docs, organize stakeholders, and get alignment; he suggests the PM-less structure may have succeeded despite that gap, not because PM work is unnecessary. He sees innovation in the merge of crafts—design plus PM, or design plus engineering—because one person operating across two functions can produce insights that specialists miss.

As Snap grew to hundreds and then thousands of engineers, projects became six-month or one-year investments, so Gaurav helped create design engineering: people who could design UX, build it, and launch it. The small team prototyped ideas inside Snapchat, tested them in places such as Australia or a couple of high schools, and gathered real-world data before committing production engineers. It also created internal virality: prototypes spread from engineers to managers to VPs to Evan, producing instant alignment and the feeling that everyone was working on the future.

Snap did not have a user-research function for a long time—Gaurav recalls one researcher until roughly 5,000 employees and the post-IPO period—but its product leaders and CEO were steeped in how users behaved. He agrees with the host’s framing that research should update a leader’s mental model rather than dictate the next feature. At Snap, the mission itself was the filter: enable private sharing in a safe way, and reject ideas outside that scope no matter how exciting they were.

That filter helps explain why Snap did not become the company to discover short-form algorithmic video. Public sharing can enable bullying and bad behavior, which Snap was trying to avoid; Instagram Stories let users reshare someone else’s story to their followers, but Snap would not, because the original person had not consented to that audience. The mission was applied through pillar-based thinking: if an idea sat outside the mission, it was out of scope.

Our Stories, including campus stories, looked like an early TikTok and produced viral moments and strong engagement, but Snap was against algorithmic ranking of that content. A curation team reviewed every post to prevent negative behavior, which was not scalable; even with high engagement, the product died because a person could not check everything. Gaurav says this was a good thing in some ways because the mission was solving a real problem, but it also meant Snap missed a major shift.

He argues social media has split in two. Traditional social networking is sharing with friends, and the old friend-to-friend virality loop is over; the new virality comes from algorithms deciding whether a piece of content is worth showing to an arbitrary audience, as on TikTok, YouTube Shorts, and Instagram Reels. That changes how people interact and how things spread, and he thinks regulators should treat the two models differently because algorithmic distribution is partly the company’s voice.

Gaurav says a world where video, images, and audio cannot be trusted has not existed for a while, but fully generated reality could return us to a pre-evidence world of physical proof and he-said-she-said. Today’s models are almost photorealistic, but he thinks fully undifferentiable video is still a couple of years away. The implications are use-case driven rather than uniform.

Captions focuses on talking video—dialogue and monologue—rather than silent B-roll, because movies, TV, social posts, and ads are largely people talking and interacting. He says avatar-video companies often use neural rendering, a separate technology not driven by transformers and diffusion models, which requires training on individual people and may not generalize. General text-to-video models can do more but have not yet reached the frontier of expressive talking video.

The safety framework divides video into documentation and storytelling. Documentation includes personal memories and journalistic or historical records, and Gaurav sees no benefit to AI-generated fake versions—only negative outcomes—so Captions wants to make that use case difficult. Storytelling includes ads, social posts, TV, and movies; nobody thinks the Geico gecko is really selling insurance, so enabling more people to tell stories is positive and is where Captions wants to make generation easy.

ByteDance’s new model, which can turn a photo into an expressive talking person, falls squarely in Captions’ talking-video territory and shows a large company using a true large diffusion model rather than simpler avatar technology. Gaurav also says AI video is already changing marketing: about a year ago ads made with AI video drew “this is fake” comments, but the technology improved enough that comments stopped and performance could beat filming a person because the team could generate 30 or 40 variants and let the winner emerge. Localization became easier too, since translating a winning creative with AI performed almost as well as the original language version, and he expects a future social network where content is generated and tailored to each viewer.

Captions began by building the Captions app in two days and launching it; it immediately took off, reaching the top of the app store and about 600 videos a day, which shocked Gaurav because he had built it on a weekend. Instead of treating that as product-market fit, the team got distracted and spent at least a year to a year and a half exploring social networks and other products. He says they should have been working on Captions.

The wake-up call came when he opened the personal account that held Captions about a year and a half in and saw $500,000 in revenue growing on its own, with no employees, releases, bug fixes, or customer support—and roughly 2,000 unanswered support tickets. He realized the product worked; the team reset and returned to Captions. Once they started releasing features, what had looked like a vertical line became the new horizontal baseline, and the new growth line was even steeper.

The original logic was to build a creation tool before a social network, and captions were chosen because speech-to-text was already solved. Then GPT and related advances solved many formerly unsolved problems very quickly, making the timing almost perfect. The story also illustrates how hard social networks are to create and kill: Captions removed its old social app from the app store, yet thousands of people still used and posted on it, and Gaurav points to X as another example of network effects surviving brand, team, and URL changes.

In the lightning round, Gaurav says he deliberately does not read books, choosing instead to build his skill in listening and watching because he thinks that is the future. He enjoys Silo and Severance, uses Linear and Superhuman daily, and keeps coming back to a motto he learned by hearing himself repeat it: to win you have to be the best, and the easiest way to be the best is to be first. He also names the ByteDance talking-video demo—including a talking broccoli—as a recent wild AI-video moment.