Meeting task clips
Staff often leave a meeting with the bosses unsure what was asked of them. This skill turns the meeting into one page they can act on, plus clips to replay the exact instruction in the boss's own words.
The user is not technical. They send a video or a Google Drive link and expect an answer. Never ask them to run a command, export a file, or install something. Do every step yourself and only come back to them for things only they can do (changing a file's sharing setting, telling you who is who).
Scripts live in the scripts/ folder next to this file. Call them by full path, e.g.
python3 <this-skill-folder>/scripts/parse_transcript.py. On Windows use python if
python3 is not found.
JOB 1 Get the meeting -> JOB 2 Get the words -> JOB 3 Understand -> JOB 4 Clip -> JOB 5 Report
link or file transcript or tasks, Q&A, one clip page + answer
fetch_input.py speech-to-text topics, who per task in the chat
parse_transcript.py (your job) cut_clips.py render_report.py
transcribe.py grab_frames.py
Job 1: Get the meeting
Make a working folder meetings/<date>-<short-title>/ in the current directory and put
everything there. What the user gave you decides what to do:
| The user gave | Do this |
|---|---|
| A video or audio file | Use it where it is. |
| A transcript file or pasted text | Save pasted text as transcript.txt. |
| A Google Drive / Docs link | If you have a Google Drive tool or connector, use it. Otherwise python3 scripts/fetch_input.py "<link>" --out <folder>. |
| Both a recording and a transcript | Best case: use both. |
fetch_input.py exit codes and what to tell the user, in plain words:
4(private file): "I can't open that link. In Google Drive, click Share, set General access to Anyone with the link, and send it again. Or download the video and send me the file."3(folder link): "That link is a folder. Right-click the video itself, choose Share > Copy link, and send me that."
If they only sent a recording, check once whether a transcript exists ("Was transcription turned on in the meeting? If yes, send me that link too; names will be more accurate") but do not wait for an answer. Carry on with the video.
Roles: if the user did not say who the boss is and who they are, infer it (the boss gives instructions and approves) and state in your answer that you inferred it.
Job 2: Get the words
Transcript available (the more accurate path, speaker names included):
python3 scripts/parse_transcript.py transcript.txt -o segments.json
Handles Google Meet transcript Docs exported as text, .vtt, .srt, .sbv, and plain
Name: text lines.
Only a recording:
python3 scripts/transcribe.py recording.mp4 -o segments.json # add --language id if known
This listens to the recording on the user's own computer. The first run sets itself up
and downloads a speech model, which takes a few minutes; later runs take roughly 5 to 10
minutes per meeting hour. Tell the user it is working and roughly how long, then let it
run. --model medium is more accurate for noisy audio or mixed languages and slower;
--model base is faster.
Both produce segments.json: a list of {id, start, end, speaker, text} with times in
seconds. Read the stderr line to see the language, count, and speakers.
Job 3: Understand the meeting (your job, not a script's)
Read all of segments.json. Long meetings can be 1000+ segments; read in chunks, do not
skip. Then write meeting.json following references/output-schema.md.
Who is speaking, when the words came from the video. Speech-to-text gives no names. Recover them from three sources, in this order of trust:
- Names used in the conversation. "Rina, can you redesign…" followed by "Okay" means the reply is Rina's and the request came from whoever runs the meeting.
- The picture. Google Meet labels each participant tile and highlights the active
speaker. Grab frames at the start of the moments that matter (each task request, each
answer; 10 to 30 frames is plenty, not every segment) and look at them:
python3 scripts/grab_frames.py recording.mp4 --at 61.9 384 570 --out frames - What the user told you about who is who.
Where you still cannot tell, write "Unidentified speaker" and list it under
open_points. A wrong name is worse than an honest blank.
Use the picture for context too. If a slide, spreadsheet or design was on screen
while a task was explained, a frame tells you what "this page" or "that number" meant.
Put what you learn into what_is_needed.
What to extract, in order of importance:
- Tasks. Any moment someone with authority asks someone to do something ("can
you…", "please…", "I need…", "by Friday…", "who can take…" followed by a volunteer).
Record the task in one clear sentence,
assignee,assigner, deadline if said,what_is_needed(details, constraints and acceptance criteria, including ones said minutes before or after the request), the exactquote, andstart/endcovering the whole explanation. A good clip runs 20 to 120 seconds: begin where the boss starts giving context, end when the assignee acknowledges or the topic changes. - Questions and answers. Every real question, who asked, who answered, the answer. Unanswered questions are open points.
- Topics in order, each with a short summary and what each speaker contributed.
- Decisions stated as final.
- Open points: no clear owner, vague deadline ("soon"), unidentified speaker.
Assignment rules of thumb, because a wrong assignee is the costliest mistake:
- "You" means the person who spoke just before, unless a name was used.
- A task given to a team belongs to the team; note who from it was present.
- Boss asks, someone volunteers: assignee is the volunteer, assigner is the boss.
- If you had to guess, set
confidence: "low"with a note. The report shows a "please confirm" flag.
Write in the language of the meeting unless the user asks otherwise, in plain words.
Job 4: Cut the clips (when there is a recording)
python3 scripts/cut_clips.py meeting.json --video recording.mp4 --segments segments.json --out clips
One clip per task, question and decision, with 3 seconds of padding, a subtitle file
beside each, and the clip names written back into meeting.json. It is a fast stream
copy, seconds even for a two-hour recording. --audio-only makes small files for chat
apps, --pad 8 adds context if a clip starts mid-sentence, --burn-subtitles draws
the words into the picture.
Timestamps from a Google Meet transcript Doc are coarse (one per block of several
minutes), so those clips can be off by a few seconds. Timestamps from transcribe.py
are accurate to the sentence. When you have both a transcript Doc and the recording and
precise clips matter, run transcribe.py as well and take the times from it, keeping
the names from the Doc.
Job 5: Report
python3 scripts/render_report.py meeting.json -o MEETING_SUMMARY.md --clips-dir clips
Writes MEETING_SUMMARY.html (a page with the clip players built in; this is the one
to give the user) and MEETING_SUMMARY.md. Read it once as the staff member would and
fix anything unclear in meeting.json, then re-render.
Then answer in the chat itself. The user should not have to open a file to learn
their tasks. Reply with, in this order: one or two sentences on what the boss wants,
the user's own tasks with deadline and the key details, tasks given to others in one
line each, anything they need to confirm, and last the path to
MEETING_SUMMARY.html and the clips folder. If your environment can display files or
video, show the summary page.
Output layout
meetings/<date>-<title>/
├── recording.mp4 / transcript.txt what the user sent
├── segments.json Job 2
├── meeting.json Job 3
├── frames/ Job 3, only if you looked at the video
├── clips/ Job 4
├── MEETING_SUMMARY.html Job 5, open this
└── MEETING_SUMMARY.md
When something is missing
- ffmpeg not installed (needed for clips, frames and listening to video): install it
for the user (
brew install ffmpeg,winget install ffmpeg,sudo apt install ffmpeg), telling them in one line what you are doing. If you cannot, and there is a transcript, still deliver the summary with timestamps and say clips were skipped. - Transcript without timestamps and no recording: summary only, no clips.
- Very long or poor audio: say so, suggest
--model medium, and mark doubtful passages as low confidence rather than inventing words. - Several meetings: one folder per meeting, then offer a combined task list.
references/output-schema.md has the exact meeting.json fields.
references/transcript-formats.md explains the Google Meet file formats.
assets/example/ shows a finished result.