Shipping fast and winning in AI: lessons from Captions and Snap

Executive overview

Most AI startups drown in roadmap options while competitors catch up. Captions runs on two disciplines: every engineer ships one marketable feature per week, and the roadmap is split into a public list (what users asked for) and a secret list (what nobody asked for yet).

The public roadmap keeps the lights on. The secret roadmap is where the wins come from.

The easiest way to be the best is to be first.

Shipping one marketable feature per week

• A marketable feature is one a user might subscribe or pay for on its own — not a table-stakes improvement.
• Cut scope ruthlessly until removing anything more would make the product useless; never cut quality.
• Ship that stripped-down version, watch complaints, then build exactly what users flag next.
• User complaints are confirmation of interest — silence is the red flag.
• Volume of shipped features creates noise-above-the-noise; teams that ship weekly stay top of mind.

The two-roadmap system

• The public roadmap is every feature users have requested — also known by every competitor.
• Prioritising the public roadmap well earns incremental wins but won't change competitive position.
• The secret roadmap contains ideas nobody asked for; users shown them might say "I don't need this."
• Secret roadmap ideas come from quarterly company-wide brainstorms — engineering, recruiting, marketing all contribute.
• The team votes, ranks, and hands feasibility analysis to the product team.
• Eye contact correction (shifting recorded eyes to face the camera) was a secret-roadmap idea; it became viral in every language and was widely copied.

Technical debt as strategic leverage

• Startups should take on technical debt deliberately — it is how a small team outpaces a large one.
• Frame it as financial debt: debt creates leverage; taking on a mortgage lets you buy something you couldn't otherwise afford.
• Every piece of debt charges daily interest: roughly 1–2% of engineering time per debt item.
• When interest compounds to 80–90%, the team is only keeping the lights on — that is the failure mode.
• Use Q4 as an infrastructure quarter to pay down accumulated debt before the next growth cycle.
• If the company fails, the future engineer who would have fixed the debt is never hired — so the debt never mattered anyway.

How Snap's design-led structure worked

• Snap had roughly 10–12 designers at 5,000–6,000 employees; no formal PMs until monetisation became central.
• Designers functioned as PM + IC designer simultaneously — writing docs, owning roadmaps, managing shipping schedules.
• No designer had direct reports; everyone designed everything themselves.
• Because all UI changes required a designer, and designers reported directly to the CEO, Evan Spiegel maintained full product context across the entire app.
• That single-mind awareness let him catch interdependencies and push non-incremental changes that others disagreed with — and were right about anyway.
• Consumer social products may structurally require one person who holds the entire product model in their head.

Design engineering: prototyping at startup speed inside a big company

• As Snap scaled past 500 engineers, every project became a 6–12 month commitment — wrong bets were expensive.
• A small design-engineering team (people who could design, build, and ship) created working prototypes baked into the live Snapchat app.
• Prototypes ran in limited geos (e.g. Australia, a few high schools) to collect real behavioural data before full build.
• Sharing internal prototypes created internal virality: engineers shared builds, VPs heard about them, Evan heard last — instant company-wide alignment without stakeholder meetings.
• The same logic applies at Captions today: ship in one week and get user data, which is better than a prototype anyway.

Functions should overlap

• Snap's unique insight was merging designer + PM into one role; the real principle is every function should deeply understand every other function.
• PMs at Captions are expected to own all the way through to marketing, because marketing is just extending the product's surface area.
• A Facebook ad is a button placed inside another app; the PM should own the full funnel from that first click.
• Engineers should understand product; designers should understand engineering; PMs should understand growth.

AI video: where things are and where they're going

• Text-to-video today is mostly silent B-roll; the unsolved frontier is talking video — dialogue and monologue with matching audio.
• Two separate technology tracks exist: neural rendering (avatar-style, person-specific, not LLM-driven) and large diffusion models (generalizable, scalable, where the real progress is happening).
• Photorealistic AI video is a few centimetres away; fully indistinguishable talking video is roughly two years out.
• ByteDance's OmniHuman is the first major example of a large diffusion model producing expressive dialogue video.
• The Will Smith spaghetti video represents the state of the art from 18–24 months ago — the progress delta is massive.

Safety framework for AI-generated video

• Video divides into documentation (recording reality — personal memories, journalism) and storytelling (ads, social media, entertainment).
• Generating fake documentation video has no upside and creates real harm; generating storytelling video is net positive.
• Product design should make the documentation use case hard and the storytelling use case easy — that tension is the real design challenge.

AI video is flooding the marketing channel

• One year ago, AI-generated video ads attracted comments like "this is so fake."
• That threshold has been crossed; AI ads now outperform single human-recorded creative because you can generate 30–40 variants and let the winner emerge.
• Localisation with AI (translating winning creative into other languages) now performs nearly as well as the original.
• AI-generated content will flood performance marketing channels — wherever there are dollars to be saved, the economics are inevitable.

Captions' origin: ignoring product-market fit for 18 months

• The app was built in a weekend, launched, hit the top of the App Store immediately, and was then largely ignored by the founders.
• The team chased social-network ideas instead; captions kept generating revenue with no releases, no support responses (2,000 open tickets), no marketing.
• 18 months later, the founders found $500k of organic revenue sitting in the account, growing on its own.
• Resuming development caused growth so steep that the prior growth curve flattened to horizontal by comparison.
• The lesson: when a product grows without you, that is the signal — go work on that thing.