How Flamme grows

Carrot: when the daily question goes live in the morning, send a notification ('Hey, come back, answer your question'). Stick: later in the day, if still unanswered, send a notification that you'll lose your streak and your partner will be sad/missing out. The stakes are tied to the partner's happiness, not the streak itself.

Noticed users came back not for the AI coach but for a side feature: a daily prompt couples answer together. Extracted it and made it the mainstay, shown big on the home page ('daily discovery question'), designed as a daily habit trigger based on BJ Fogg's habit research. Took ~8 months of iteration.

Compared formats by view-to-download conversion: memes and slideshows 0.05-0.2%, reaction hook followed by an app clip 0.3-0.5%, videos where the app in use is the main subject 0.5-1%. Stopped making meme videos even though one got 300K+ views; show the product in action as bluntly as possible.

After a 0.8% D30 date-planning MVP (a once-in-a-while tool, like booking an Airbnb), relaunched in December 2022 as Flamme AI, an AI relationship coach (ChatGPT wrapper) website, riding ChatGPT's release; it was pitched as the first AI for relationships and got TechCrunch and LA Times coverage.

Treat TikTok as a volume game: take formats that work and replicate them across a couple dozen accounts (20-30 for Flamme). Numbered iPhones (many iPhone 8s; 10 with the team now, 3-4 accounts each). Each phone is factory reset with no leftover fingerprints, gets its own Apple ID and a paid proxy/IP client to reach the US For You feed from outside the US. Natively created accounts are safest (a Europe-created account was not relocated because relocating costs a lot). Accounts aren't named Flamme and aren't shared on X, since formats stop working once everyone copies them.

To download the paid proxy/IP client once, logged into every phone with the same Apple ID. TikTok fingerprints Apple IDs, linked all accounts as one device regardless of different IPs, and one flagged account hurt the rest.

TikTok slideshows titled like 'five uncomfortable but healthy questions to ask your partner', each slide a sample question from the app, watermarked 'Flamme: The Couples App'. Designed for shares and saves.

Launched the first version (Sparks, a date planner) in October 2022 while being featured on TechCrunch as a Startup Battlefield company; no TikTok or paid. Follow-up AI relaunch got TechCrunch (Valentine's Day) and LA Times coverage.

Under 300 views: account problem (shadow-flagged or banned, setup wrong), fix the account. 300-1,000 views: content problem, try new formats and hooks. 1,000-5,000 views: keep the same format and type, change the copy slightly until it crosses 100K views.

Each playbook is a ~10-page PDF the team feeds to AI. Section A: product and brand context (name, category, what it is in human terms, core features, target audience, psychology). Section B: winning script references (own and others' winning content in the category, with 'what works, what to avoid, do not repeat them') plus screenshots of top slideshows. Section C: master prompt with a system role ('You are a viral TikTok question slideshow copy engine for a couples app called Flamme'), context and output requirements (title slide, set categories, a Nano Banana image prompt). Takes the founder a full working day to build; he tests the output before handing it over. GPT writes the image prompt; Nano Banana Pro makes realistic vertical images. Winners are added by hand to the master prompt (AI edits hallucinated), creating new playbooks.

A manual Google/Excel sheet with one row per phone/account: account status and health, whether it's online, content type, notes, posts per day, plus daily views and likes; performance checked on day 2. A 1-hour team meeting every morning reviews it. Winners go back into the AI playbook; unhealthy accounts get recovery attempts, then are ditched and replaced. He avoids auto-tracking tools for own accounts because they make you stop looking at each piece (tools are fine for UGC at scale).

V (virality): the format got 100K+ views on accounts with under 5,000 followers and is trending across other accounts. S (scalability): it can be recreated consistently with AI tools across 5-10 accounts several times a day (Sora, theinfluencer.ai, Nano Banana Pro; Veo videos were flagged as AI). A talking video you must record by hand fails this. C (convertibility): comments ask 'What app is this? Where can I get this?'; if not, it gets views but won't convert.

If you haven't answered the daily question, you cannot see your partner's answer. Each partner answering pulls the other to answer, creating a loop.

An Instagram-native founder spent 4 hours every day for 2 months scrolling TikTok, studying the language, content types and hooks, and why some posts (including their own) got traction and others didn't. Also learned from Consumer Club members.

Copied Couple Joy, one of the first couples apps to put a widget on the lock screen (and base its social media strategy on it): widgets showing the distance between partners (valuable for long-distance couples) and the number of days together, updated dynamically so a large widget reminds users of the app every time they unlock the phone. Discovered late, after notifications.

With ~50% D30 retention but small absolute numbers, pitched General Catalyst, a16z, Sequoia and a16z Speedrun. Investors liked retention, then asked how the product would be distributed; vague answers about influencers ended talks after one or two rounds of calls. Guest now advises unlocking distribution before raising.

For content featuring the app directly, optimize for likes (a like means it resonated; million-view videos with low like ratios converted poorly). For question slideshows, the best performers had high share and save ratios (people send questions to their partner or save them for later). Across formats, total watch time mattered more than the first 3-second hook, e.g. curating memes to raise watch time.

Flamme was completely free through 2023. Near the end of 2023, almost out of money and unable to raise, added monetization. Mistake: calculated conversion as all paying users (new subscribers plus renewals) divided by incoming users, which showed 15% in month one and made them think they could scale.