Find all claude sessions (claude CLI and from claude code desktop which are stored in
a different location) to find the highest-leverage improvements to my setup and usage
patterns. Use subagents to pull raw signals from the
transcripts; you cluster them across sessions and decide, per cluster, whether
it wants a new skill, an automation, a fix, or nothing - weighing recurrence
against build cost. Is there anything I should be doing more of? Are there are
particular prompts that could bring a lot of value to my projects?
Write candidates to reflection-notes.md with the evidence
(which sessions, how often) so it's my record afterward.
This is diagnosis only - don't build or edit anything; the deliverable is your
ranked assessment, most leverage first. Only propose a skill for something that
actually recurs, and cite the sessions behind each call.