What they did
Mekonnen, a backend and AI engineer, wanted a Telegram channel about AI coding tools that ran with no server, no per-token bill and no approval step. He set up Claude Code routines that start a fresh cloud session at 07:00, 12:00 and 16:00 UTC, plus a Sunday roundup with a poll and a monthly pricing cheat sheet. Each run clones a GitHub repo and follows a PROMPT.md file: it scans vendor blogs, changelogs, release trackers, GitHub and web search, scores stories on developer impact, freshness, official sourcing and reach, and writes the best one as a small JSON post. A plain Python script, send.py, then checks the post (length limits, allowed emoji and hashtags, no links, no repeats of an already posted story), finds or draws an image card and sends it through the Telegram Bot API. The repo holds the rules, the list of past posts and a log of every run, and each run commits back to it.
How it went
Day one ended with seven posts: news, tips, an explainer and the pricing sheet, all written by the agent. Three things broke. The first run could not reach github.blog until he widened the cloud environment's network access. One post went out without an image because the image library was broken in the cloud Python and GitHub blocked a page fetch. An explainer packed five numbered tips into one paragraph, which he fixed by hand, the only manual edit so far. Each fix went into send.py as a check with a test rather than into the prompt; the script now has about 90 tests.
Worth knowing
When the model writes a post that breaks a rule, the script refuses it and returns the error, so the run has to fix the post before anything is sent. Routine runs per day are capped by the Claude plan, so he keeps to three posts plus the weekly and monthly jobs and tests locally with a dry-run flag. This is an account of the first day live, not of months of running.