kibitzer-check-brainstormlisted

# Kibitzer Check Brainstorm kibitzer (`github.com/tstapler/kibitzer`) ships a small set of built-in checks (`src/primitive_obsession.rs`, and whatever else is registered in `src/check.rs`). This skill looks for evidence of *new* check ideas — patterns of agent mistakes, sloppy edits, or missed problems that show up repeatedly across this machine's Claude Code session history — independent of whether kibitzer currently has a check for them. Contrast with `kibitzer-sample-review`, which only reviews firings of checks that already exist. The output is a set of candidate checks, each backed by cited real transcript evidence — not a hypothetical wishlist. An idea with zero observed occurrences is a guess, not a finding; note it separately if it's still worth recording. ## Step 1 — Survey what mistake patterns actually recur Don't grep for a specific keyword up front — that presupposes the category of mistake. Instead sample broadly across sessions and look at what assistant turns get corrected, reverted, or flagged by the user. Useful raw signals, per transcript file in `~/.claude/projects/*/*.jsonl`: ```bash # User turns that read as corrections — cheap first pass, high false-positive rate, # but a good way to find candidate sessions worth reading in full. for f in ~/.claude/projects/*/*.jsonl; do jq -r 'select(.type=="user") | .message.content | if type=="string" then . else (.[]? | .text? // empty) end' "$f" 2>/dev/null \ | grep -iE "no,? (that|don'\''t)|revert|