Wend: an AI copilot that learns a city guide's voice and drafts their weekend digest

Every Friday, the same job: open the tabs, copy the listings, fight the formatting, second-guess what is actually worth anyone's time. The assembling is the easy part to automate, and the wrong part to lead with. A generic AI digest reads like a press release, and for a one-person channel the voice is the business. So Wend does not start with an event list. It starts by reading the operator's existing posts and distilling a six-part profile: voice, humor, taste, positioning, signature, and formatting. The taste section is the lever, fed back into event selection as criteria, so the digest picks what that channel would pick and writes it the way they write. AI can generate a hundred viable options in the time it once took to make one, and still not tell you which lands. That judgment is the moat, and Wend learns it from the work the guide already did.

The unglamorous half is getting the raw posts and events in at all. Three platforms, three different problems. Telegram has no open API for this, so Wend reads the public channel feed and pages backward to about a hundred posts. Instagram refuses the usual scraper outright, so those captions come through a separate one. Everything else runs through Firecrawl, which starts on a cheap proxy and climbs to a residential anti-bot tier only on the pages that actually block, then reports what it missed instead of dropping a source in silence. The listings arrive in Ukrainian, so the model transliterates every name to Latin and returns an empty field rather than inventing a placeholder. The boring correctness, dedup and null handling and honest coverage, is what makes the output safe to paste straight into a channel.