feedback-watch

name	feedback-watch
description	Watch for UI feedback annotations from the browser via the MCP bridge. Processes incoming annotations, makes code changes, and responds back to the browser. Runs in a loop until the user stops.

Wait for UI feedback from the browser, process annotations, make changes, and respond back. Continue listening until the user says to stop.

<trigger_conditions> Activate when ANY of these are true:

• User says "watch for feedback", "wait for annotations", "listen for feedback"
• User says "start feedback loop" or "feedback mode"
• After running /add-feedback, if the MCP bridge is available </trigger_conditions>

<do_not_trigger> Do NOT activate when:

• The feedback_watch MCP tool is not available
• User is asking about feedback in a general sense (not the UI annotation system) </do_not_trigger>

1. **Start watching**: Call the `feedback_watch` MCP tool (timeout: 300s) - Tell the user: "Watching for feedback from the browser. Annotate elements on the page and click 'Send to Claude'."

2. Process feedback: When feedback arrives:

   • Parse the annotations (section, label, selector, HTML, feedback text)
   • Read the source file from the feedback entry
   • For each annotation, determine what change is needed based on the feedback text
   • Make the requested code changes to the source file
   • Summarize what was changed

3. Respond to browser: Call feedback_respond MCP tool with:

   • The feedback_id from the received feedback
   • A message summarizing what changes were made

4. Continue loop: Call feedback_watch again to wait for more feedback

   • If user says "stop", "done", "exit", or "quit", stop the loop
   • Otherwise, continue processing feedback as it arrives

5. Cleanup: When stopping, tell the user: "Stopped watching for feedback. Your changes have been saved."

<success_criteria>

• Feedback received and processed within the watch timeout
• Source file changes match the annotation requests
• Response sent back to browser via feedback_respond
• Loop continues until user explicitly stops </success_criteria>