| name | chronicle-screenpipe |
| description | Record a meeting from the computer with screenpipe and turn it into a raw markdown transcript routed into the vault. Verb-routed — `start` launches a scoped recording (AUDIO-ONLY by default, add vision for slides), `end` stops it and extracts the transcript, `extract` pulls a past time window from the existing DB without recording. Hands the raw .md to /note to file a clean meeting note. Use when the user says /screenpipe, "record this meeting", "start recording", "stop recording", "transcribe the meeting", "记录会议". |
| created | 2026-06-28T00:22 |
| updated | 2026-06-28T00:22 |
/chronicle-screenpipe — record a meeting → raw transcript → vault note
Thin wrapper over a locally-installed screenpipe. screenpipe
is the capture engine (mic + system audio, optionally screen+OCR); this skill scopes
ONE meeting, extracts it from ~/.screenpipe/db.sqlite, and routes the result.
Routing — pick the verb from the user's words
| User says | Verb | Action |
|---|
| "start recording", "record this meeting", "/screenpipe start" | start | bin/sp-start.sh <slug> |
| "stop", "end the meeting", "/screenpipe end" | end | bin/sp-end.sh → then route |
| "transcribe yesterday's call", "pull the 2-3pm meeting" | extract | bin/sp-extract.py --date/--start/--end → then route |
Derive <slug> from the meeting topic: 2–4 lowercase hyphenated words
(e.g. a1c-team-sync, golden-phd-chat).
start — launch a scoped recording
bin/sp-start.sh <slug>
bin/sp-start.sh <slug> --with-vision
- Audio-only is the default (
screenpipe --disable-vision): no screen capture,
no OCR, smaller DB, fewer privacy concerns. This is what you want for most meetings.
- Records a timezone-free row-id watermark so
end knows exactly what is new.
- Refuses if a meeting is already active. One meeting at a time.
- ⚠️ Consent: this captures other people's audio. Confirm everyone is okay
being recorded before starting — same obligation as any recorder.
end — stop and extract
bin/sp-end.sh
bin/sp-end.sh --keep-running
Writes the raw transcript to
_WorkSpace/screenpipe-raw/<YYMMDD-HHMM>-<slug>.md and prints its path.
extract — a past window, no recording
For meetings already in the DB (e.g. screenpipe was running 24/7):
bin/sp-extract.py --db ~/.screenpipe/db.sqlite --out <path.md> --slug <slug> \
--date 2026-06-05
--start "2026-06-05T20:00:00" --end "...21:00:00"
route — file it into the vault (always the last step)
After end/extract produce the raw .md, hand it to /note:
it logs to the diary, keeps the raw file in _WorkSpace, and writes a clean
structured meeting note into 1-EVENT-SPACE linking the raw transcript.
Do NOT hand-author the meeting note here — that is /note's job.
What the raw markdown contains
- Header — slug, mode, time span, turn/frame counts.
- Speakers — ranked by airtime, capped at 8 (+ collapsed tail).
- Transcript — consecutive same-speaker lines merged; mic/system
double-captures deduped.
- Screens shown — only in
audio+vision mode: distinct screens deduped
by OCR similarity, with thumbnails when screenpipe saved a snapshot.
Known limits (set expectations — don't oversell)
- Diarization is noisy. screenpipe over-fragments speakers (one person → many
speaker_ids). The legend ranks by airtime and collapses the tail, but speaker
labels still need human verification/renaming in the meeting note.
- Transcription is Whisper-grade and degrades on heavy code-switching
(e.g. Mandarin+English) and crosstalk. Treat the transcript as a searchable
rough record, not a verbatim minute.
- System audio must be captured for the other participants to appear
(
is_input_device=0). If only your mic shows up, screenpipe's audio loopback
isn't configured.
Files
bin/sp-start.sh launch scoped recording, write watermark marker
bin/sp-end.sh stop + extract since watermark
bin/sp-extract.py the extractor (transcript merge, dedup, frame selection)
