AI Workflow
End-to-End Design
to Scale User Growth
I turned content ops into a product system: reusable formats, hard brand rules, an agent that builds in Figma, and a human gate before publish.
FAVORS has no marketing team, just two founders. Claude plans and builds carousels in Figma. Buffer publishes across three TikTok accounts. I keep taste and release control. One idea becomes a finished carousel in five minutes.
Production dropped from two hours to five minutes. Daily new users went from about 30 to around 75. In October, FAVORS hit #7 in Social Networking on the Swiss App Store.
Role: Co-Founder · Product, Design, Growth & AI systems
System: Claude agent · Figma runtime · Buffer distribution · RevenueCat signal
faster production, 2 h to 5 min
Problem and hypothesis
Every post cost about two hours. Two founders could not ship product and feed several channels every day. Ideas were never the bottleneck. Production, consistency, and learning speed were.
Hypothesis: if content runs as a system with reusable formats and hard rules, an agent can own production, and we can find the formats that convert without a marketing hire.
The agentic system
Connecting a model to a tool is the easy part. The product is the system around it.
Old: handmade carousels, inconsistent output, slow learning.
New: format library, Claude skills as brand law, Figma as the agent runtime, Buffer for distribution, human gate before publish.
Claude plans each story and builds it straight into the file: hook, content, CTA. I chose Figma over a template API so every carousel stays editable and one file serves craft and automation.
Format library
Tips, mistakes, founder stories, product walkthroughs and short conversion posts. Each format has one job and the same narrative arc.
That makes the agent reliable and the learning comparable. Formats are the unit of experiment, so I can compare what works and double down fast.
Built to feel like UGC
People trust people. In dating, polish often kills belief, so the design goal was the opposite of a brand deck: every slide should feel like a friend’s post.
Those rules do not live in one-off prompts. They live in Claude skills: lowercase Swiss German, slide density, type scale, CTA placement. The agent inherits brand judgment on every run.
Human in the loop
The agent owns speed and options. I own judgment: approve the story, check the visual, release the post.
No auto-publish. Automate repetition. Keep taste as a product gate. That is what keeps the system brand-safe while it scales output.
Results
Once production stopped being the bottleneck, growth could compound. Daily new users moved from about 30 to around 75 in the pre/post window measured with RevenueCat. Best day: 359 new users.
Why it matters: cadence became sustainable, format learning got faster, and channel strategy became part of the system, not a weekly debate.
#7
FAVORS hit #7 among the top free apps on the Swiss App Store in October 2026, right between Discord and WeChat. That ranking is the outside signal. Inside the company, content stopped being handmade and started running as a system.
in Social Networking
FAVORS hit #7 among the top free apps on the Swiss App Store in October 2026, right between Discord and WeChat. That ranking is the outside signal. Inside the company, content stopped being handmade and started running as a system.
more views per post on creator accounts
People follow people. Over 30 days (21 Sep to 21 Oct 2026, measured in Buffer), our two creator accounts averaged about 4,200 views per post versus about 1,060 on the brand account. So the system encodes a channel strategy: new formats ship in first person on creator accounts, while brand stays on product and conversion.
Hypothesis check
Production: confirmed. One carousel now takes five minutes instead of two hours, without a marketing hire.
Converting formats: first signal. First person formats on creator accounts get 4x the views of brand posts, so they get more variants.
What’s next
Close the loop. Tag posts by format and hook, feed views and installs back in, then let performance decide which formats get more variants and which get killed. Same skill architecture across markets, with a human gate per market.
Key learnings
Systems beat tool hooks. Design file = agent runtime. Formats are experiments.
Let agents produce.Keep judgment human.