plugins/live-transcribe/skills/live-transcribe-read/SKILL.md
Read the latest real-time transcription. Use when user asks to see, read, or show a transcription that was captured via live-transcribe. Triggers on: 'תקריא תמלול', 'מה תמללתי', 'התמלול האחרון', 'show transcription', 'what did I say', 'read the transcript', 'מה נכתב בתמלול', 'תראה לי את התמלול'. Also use when user references transcription content without being explicit — e.g. 'summarize what I said', 'translate the transcription'.
npx skillsauth add aviz85/claude-skills-library live-transcribe-readInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Read the most recent transcription file from a live-transcribe session.
The real-time transcript lives in the log file, not the .txt file. The log updates continuously during recording.
# Primary source — real-time log (updates live during recording)
cat /tmp/realtime-transcribe.log 2>/dev/null
# Extract only the spoken text (strip metadata lines)
grep '^\[committed\]' /tmp/realtime-transcribe.log 2>/dev/null | sed 's/^\[committed\] //'
The .txt file (/tmp/transcribe-*.txt) may contain a final dump but does NOT update in real-time. Always prefer the log file.
if test -f /tmp/realtime-transcribe.pid && kill -0 $(cat /tmp/realtime-transcribe.pid) 2>/dev/null; then
echo "ACTIVE — file is still being updated"
else
echo "FINISHED — file is complete"
fi
Use grep to extract committed text from the log, stripping [committed] prefixes. Skip the first JSON status line and [session started] markers.
The user may ask to do things with the text beyond just reading it:
development
The 10x10 method — generate breadth, then converge with human judgment. Use whenever a single AI output won't nail it and quality matters (design, copy, naming, posters, messaging, strategy options, code approaches), OR when the user says '10x10', 'ten by ten', 'give me 10 options', 'show me variations', or asks to refine/tighten an output instead of round-after-round corrections.
development
The 10x10 method — generate breadth, then converge with human judgment. Use whenever a single AI output won't nail it and quality matters (design, copy, naming, posters, messaging, strategy options, code approaches), OR when the user says '10x10', 'ten by ten', 'give me 10 options', 'show me variations', or asks to refine/tighten an output instead of round-after-round corrections.
development
Search across all Claude Code conversation history (JSONL files) across all projects.
development
Deep code audit that detects misleading patterns — fake tests, mock abuse, shallow health checks, overly optimistic error handling, hidden debt. Produces a structured report with findings AND actionable recommendations. Use when code looks green but smells wrong.