How Layers Runs Our Growth Loop

Audience and growth hypothesis

PoquitoTalk helps English-speaking expats, off-grid homeowners and travelers in Bocas del Toro, Panama, send polite, natural voice notes to local service providers without struggling through broken Spanish that leaves everyone guessing. They depend on boat captains, plumbers, electricians and solar techs, and a formal translated text often goes unanswered, whether the water pump just died or a tourist needs a boat to Bastimentos today. Growth hypothesis: the island's community chats ask the same questions every day ("Who can recommend an electrician?", "Does the ATM have cash?", "Which captain can take us?"), so practical island tips earn organic reach, and every voice note a user sends shows PoquitoTalk to a local worker and the people around them.

The loop

• Message: "When your water pump dies or your boat engine cuts out in Bocas del Toro, finding and texting local trades in formal textbook Spanish gets left on read. PoquitoTalk gives you the verified island directory and generates natural Panamanian WhatsApp audio with zero-install 2-way replies."
• Channel or product surface: the Layers autopilot on our TikTok (@dorienvibecodes), a closed loop in three parts:
  1. Auto-generation: Layers creates each clip itself from our brand assets and app screens, modelled on formats that already work: on TikTok, my own best posts (local parade videos); on Instagram, what works for the eight accounts Layers tracks (translator apps Owll and SayHi, Babbel, Duolingo, The Bocas Breeze and three local creators). It works from our App Description, which we rewrote to name a specific person, job and frustration, using Scott's formula; Layers' signal strength for the project is 88%.
  2. Posting cadence: one clip a day at 9:00 since September 14, published with no manual step. Instagram runs the same loop, with each post waiting in the review queue for my approval. Layers can't connect a Facebook page directly, so our Instagram shares each post to the PoquitoTalk Facebook page: both accounts carry the same reels.
  3. Data feedback: views, saves and engagement flow back into Layers, which reviews each post against our own account's median and points to the hooks and formats to double down on. The Layers SDK adds installs and opens from the app, so a post can eventually be tied to an install.
• Experiment: Compare the automated daily clips against each other and against our hand-made formats, using TikTok view data. Our first experiment came from a Layers opportunity on September 14: an "expat emergency repair" hook, registered before posting. Later we registered six more posts in Layers as tracked experiments, each testing one variable: hook, format, posting time, or posting frequency. Next to the loop, I also ran a handmade challenger: a bar skit that ends with a "save this" cheat sheet.
• Intended outcome: Find the hooks and formats that clearly beat our account's median, then feed them back into the daily loop. Once the winners are clear, connect posts to installs with Layers tracking links.

SDK installation and observability

The Layers Growth SDK (@layers/expo v3.3.0) is installed and reporting from the Android app:

• Project ID: prj_8272a6db-780a-462c-bd60-39b0434ccc1a (.layers/project.json)
• App ID: app_14732857b75e24a0 (src/services/layersService.ts)
• Verification check: Running layers call install_growth_measurement '{"action":"verify"}' confirms:
  • outcome: "reporting"
  • installState: "reporting"
  • blockers: [] (0 blockers)
  • observedEvents: ["app_install", "app_open", "install_referrer", "$sdk_health", "$first_open", "$app_foreground", "$app_background"]
  • First event September 14; latest reading on September 30: still reporting, last event September 29.
• What it measures: App lifecycle events (app_open, $app_foreground, $app_background), purchase events via layersService.track('purchase_success'), and a RevenueCat connection via connectRevenueCat.
• Honest limits we found and are fixing:
  • The SDK's optional storage package (@react-native-async-storage/async-storage) was missing, so every app launch counted as a new install. It is now installed and ships in v1.5.11, the next update.
  • Layers has recorded our first web revenue through Stripe (September 27), but no RevenueCat events yet: Android in-app purchases did not work in the earlier live build. The fix went live in v1.5.10, approved on September 29, so RevenueCat purchase events can now reach Layers.
  • Layers cannot yet tie an install to a specific post. We have not placed a tracking link yet (TikTok needs 1,000 followers for a clickable bio link; we have 844).
  • Installs so far: 6 on Google Play, with no paid ads. That's small on purpose: PoquitoTalk is hyper-local. The Bocas del Toro archipelago has roughly 13,000 to 16,000 permanent residents, and the full-time expat community, our core market, is an estimated 1,000 to 1,500 people. Late September is also the middle of the low season, with few tourists and many part-time residents away until November or December.

Learning and iteration

• What the data showed (TikTok Studio, September 14–27):
  • The daily autopilot clips settled at a stable baseline of about 210 views per post (most between 130 and 350).
  • One autopilot clip clearly broke out: "Don't end up in a messy situation 💩" reached 805 views, about 4× our median, but only 2 likes. A humorous, problem-avoidance hook got more views than calm product benefits, but people watched and scrolled on.
  • Hand-made comedy skits (textbook Spanish vs island Spanish, "embarazada" vs "embarrassed") landed between 177 and 322 views: in line with the baseline, not above it.
  • A practical tips post ("Cash in Bocas del Toro: 3 things nobody tells you") earned our first save: 225 views, 1 save.
  • Our TikTok is my existing personal account (@dorienvibecodes). Following Scott's advice on the Discord to connect more accounts, I created dedicated PoquitoTalk accounts on Instagram and Facebook, starting from zero followers. Layers can't connect Facebook directly, so Instagram shares every post to the Facebook page, and both carry exactly the same reels. The brand-new Facebook page turned out to be the bigger audience: 1.5K views and 1.3K viewers from September 1 to 28 (Meta Insights), about 200 people per reel, against 3 to 31 for the same reels on Instagram.
  • Instagram, honestly, is a mess. Layers created two AI influencer personas, one for TikTok and one for Instagram, and at my request it made a third. Instagram got very few views and almost no feedback, so I kept trying things, including cross-posting my TikTok videos there: what could it hurt? The result is an account with three faces and no clear style. The same reels did better on Facebook, at least in views. Our TikTok isn't uniform either, but its audience seems more forgiving. Different platforms, different audiences, different expectations: we haven't figured out Instagram yet.
  • Real footage beat the remakes. On my existing TikTok account, Layers modelled its clips on my most successful posts: real videos of local marching parades (where the jingle in our demo comes from). For the new Instagram and Facebook accounts, it drew more on what worked for the accounts it tracks. Honest observation: none of the generated remakes came close to my real parade videos. In a small, close-knit community, real local footage seems to land better than AI-generated video, while AI-made content did better in comedy skits: our best AI-generated post so far is a comedy skit with a "save this" cheat sheet at the end (see below).
  • On September 25 we posted nine clips in one day, and the later posts got only 6 and 7 views. Posting too often may have played a part: Layers' posting research (12.68M videos from 189,718 creators) shows that three posts a day bring about 3.3× the total daily reach of one, but nine is far beyond what that research covers. A better clue came on September 29: a video I had uploaded by hand from my desktop stalled at 0 views twice, then got views within minutes when re-uploaded from the TikTok phone app. The posts Layers publishes automatically never stalled. My guess is the desktop upload, possibly while I was connected to my work VPN, which can make TikTok place the post in the wrong country. So we suspect two things at once: the upload conditions, and simply posting far more than three times a day. That re-uploaded video (the bar skit, re-cut with a "Panama bar cheat sheet" ending to earn saves) now has 332 views, 7 likes, 1 comment and 1 save (2.7% engagement). Not spectacular in views, but it is the best-performing AI-generated video we have posted so far, in any genre: parade remakes, comedy skits with or without a cheat sheet, and practical tips. Compared with the 805-view breakout, it has less than half the views but 3.5 times the likes, plus a save: views and engagement tell different stories, and a useful ending (the cheat sheet) seems to turn a laugh into a like or a save.
  • A handmade challenger, next to the loop. The bar skit wasn't made by the autopilot: I made it myself, as a challenger running alongside the Layers loop, not a replacement for it. Its numbers flow back into Layers like every other post, so Layers can learn from it and carry what works into the clips it generates next. The first signals are small but promising (7 likes, a comment and a save on 332 views), so we'll keep testing this format: a short skit that ends with a practical cheat sheet people want to save.
• What Layers could confirm: Layers has returned two verdicts so far, and both are inconclusive: the September 14 "expat emergency repair" hook, and the humorous "messy situation" hook behind our 805-view breakout. On the 805-view test, Layers put it this way: "This imported post is a retrospective observation and cannot establish a causal winner." In other words, one post that took off, registered after the fact, doesn't prove the hook caused it, and Layers doesn't pretend otherwise. The other five experiments were registered after their posts went live, so Layers treats them as retrospective observations; they were still awaiting results at submission time, so the numbers above come from TikTok Studio.
• Why we held back on paid ads: As Scott (Layers) put it, "paid amplifies whichever part is already working." We had one breakout post, not a repeatable winner, so we kept paid media off and focused on the organic loop.
• What we learned: Scott shared on the Shipaton Discord that what you post explains five times more than when you post, and that 1 in 100 posts earns half of an account's total views. Layers' posting research (TikTok posting rules, 12.68M videos from 189,718 creators) puts it plainly: "When you post is noise. What you post is signal." Our small sample points the same way: the hook mattered far more than the format or the posting time. It is a small proof, not a conclusion, but it gives us a clear direction to iterate on.
• What we'd try next (the next 4 weeks, as a solo founder):
  1. Real footage vs. remakes, head to head. Film real island moments myself (the docks, a water taxi ride, the next parade) and put a POV caption on them. Post 3 of those and 3 Layers-generated clips in the same weeks, and compare views and saves.
  2. Retest the humor hook properly. The "messy situation" hook got 805 views once, but Layers rightly called it inconclusive. Run that style 3 more times, registered in Layers before posting, to see if it holds up.
  3. Register every test before it goes live, so Layers can give a real verdict instead of a retrospective one.
  4. Keep the daily autopilot, add my own posts on top: about 2 posts a day in total, manual uploads from the phone app and never over a VPN.
  5. End every tips video with a "save this" card and track saves, since saves are the strongest signal in Layers' research and our cash tips post earned the first one.
  6. Connect posts to installs: add a Layers tracking link to our website and X profile now, and to the TikTok bio once we pass 1,000 followers.
  7. No paid ads yet. Only when one format beats our median several times in a row, and the SDK shows installs following it.
  8. Figure out Instagram before restarting it. Archive the off-brand posts (reversible, and the handle and the Facebook link stay), choose one presenter, and test what Instagram's audience responds to, instead of copying TikTok.