Case in progress.
I'm still finishing this page right now: writing the copy, tidying up the screenshots and pulling the press together. Check back a little later, or take a look at the finished cases nearby.
An AI pipeline for a personal channel
Developer · 2025
The AI channel started eating more time than I wanted: gathering materials, transcribing voice notes, reading articles, assembling posts — several hours of routine a day. I decided to automate the drafting, but keep the final call for myself. That's how the bot came about: it collects materials from different sources, runs them through a chain of services, and returns a post draft. The hardest part turned out to be not the integrations, but the attempt to preserve my own voice.
The problem
Running a channel turned out to be less about writing texts and more about preparing materials. Over a day, links, voice notes, screenshots, and ideas pile up that then need to be assembled into a coherent whole.
I tried storing notes in different places and processing them once a week. In practice that just moved the same hours to another day.
So the task became automating exactly the prep stage, not the authoring itself.
The solution and the pipeline
The bot takes a link, a voice message, or a screenshot. Links are cleaned and extracted via trafilatura, voice messages go through the Whisper API. Then all the context is assembled into a single structure and sent to the Claude API for draft generation. The /go command starts the build and returns the finished text back into Telegram.
Status
The project is at the working-skeleton stage.
The basic scenario already works: I send material, run processing, get a draft. For daily use that's enough.
The most unstable part is the authorial voice. I keep rewriting the prompts: good posts come not from the model but from the context and the editing.
What's next
I want to bring the system to a state where it doesn't write posts for me but assembles a solid first version. The next step is to store the channel's context better, take past publications into account, and cut down on manual edits before publishing.