A Telegram feedback pipeline that hands testing rounds to a coding agent

Started from a real annoyance: testing a freshly launched site across desktop and mobile while dumping screenshots into Apple Notes. The session was a design conversation, not a code-writing one — the output was a set of behaviours and constraints, which only later became packages. The useful moves were mostly subtractive. Screenshot-with-caption instead of screenshot-then-note, because pairing them later is where these systems break. A tag taxonomy sorted by what the fix is rather than what the symptom looks like, which is what stops a conditional-render bug getting filed as a copy tweak. Two output formats aimed at two different readers instead of one format compromising for both. And a hard scope on where the capture agent may write, so an agent with shell access never gets loose in the source tree. The decision I keep coming back to is the one to not automate the handoff. It would have been easy to have the implementation agent watch the folder and start work when a round lands. Deliberately didn't. The gap between capture and implementation is the review gate, and closing it would have removed the only place a human looks at the whole round before code changes. What the later /feedback round command changed was the cost of crossing that gap, not the fact of it — one keystroke instead of a typed paragraph, still a keystroke. The generalization step was the surprise. Once the thing worked for one site, almost everything project-specific turned out to be a parameter — root path, site name, whether a coding-agent inbox exists at all. What was left was a skill anyone can install.