Transcribe Meeting
A recording becomes a usable brief in three tool calls. Each one produces a file
the next one reads — nothing is passed inline, because a 46-minute meeting is
~135,000 characters against a 60,000-character tool-result budget, and a summary
built from the truncated half looks finished rather than truncated.
| # |
Call |
Produces |
| 1 |
transcribe_media(file_path) |
Raw transcript at ~/.gaia/transcripts/<name>.txt. Returns transcript_path, never the text. |
| 2 |
refine_transcript(transcript_path) |
A speaker-labelled markdown transcript, already indexed. Returns refined_path, speakers, turns. |
| 3 |
summarize_document(refined_path, summary_type='detailed') |
The brief. |
Procedure
Confirm the file exists, and set expectations. If the path is not there,
say so and quote it — never describe a recording from its filename.
Measured on a 46-minute recording: ~6 min to transcribe, ~4 min to separate
the voices, ~1-2 min to name them and summarize — about 11-12 minutes, or
roughly 4x realtime. Say that before you start, not after the user has
waited.
Transcribe. transcribe_media(file_path). Pass output_path to write
elsewhere, language to skip auto-detection. The transcript is written
before anything downstream runs, so it survives a later failure; if the call
errors, report its message as given — it names the missing component and the
command that installs it.
preview is the first 1200 characters, for identifying the recording only.
low_confidence_spans and low_confidence_span_count are informational —
the transcript keeps the words as heard. Do not repair them yourself.
Refine — one call, and it is not your work to redo.
refine_transcript(transcript_path) groups the transcript by the voices
already separated from the audio, puts names to them where the conversation
supports it, writes the result and indexes it. You do not work out the
speakers yourself, do not rewrite wording, and do not write a transcript
file. Summarizing the raw transcript instead produces a brief with no owners
on the action items.
It needs the timings saved alongside the raw transcript. If it reports they
are missing, re-run transcribe_media on the original media rather than
trying to segment the text yourself — turn boundaries guessed from prose are
invented, not observed.
Report both paths, every time. State the raw transcript_path and the
refined_path in your reply — even when the user only asked for action
items, even if a later step fails. Transcription costs minutes of compute; a
user who does not know the files exist pays for it twice.
Summarize. refine_transcript already indexed the file, so go straight
to summarize_document(refined_path, summary_type='detailed'), which folds
the whole transcript forward in sections so the brief covers the entire
meeting.
Remember the outcome, not the transcript. remember a short record of
the meeting: what it was, when, who was in it, the decisions, and any action
item the user personally owes. Two or three sentences.
The transcript is already indexed — that is what detailed questions read
from. Memory is for the durable facts that should surface without being
asked, weeks later, when the user says "what did I commit to?" or a related
topic comes up in a different conversation.
Do not store the transcript, long quotes, or the full brief in memory. That
duplicates the index, crowds out everything else the user asked to be
remembered, and gets recalled in conversations it has nothing to do with.
Store nothing when the recording turned out to be something other than a
meeting.
Invite the follow-up. Close by telling the user they can ask questions
about the meeting and you will answer from the indexed transcript. They will
not discover this on their own.
Never build the brief from a fragment
- Not from
preview. 1200 characters tells you which recording this is and
nothing else — not the topic, not the participants, not the decisions, and
certainly not the action items, which cluster at the end.
- Not from
query_documents. It returns only the top few matching chunks,
so a summary built from it silently omits most of the meeting.
query_documents is the right tool for a later follow-up question about a
specific detail.
If a stage did not run, say which one and what the brief is therefore missing.
Do not present a partial pass as the finished product.
When speaker identification did not run
transcribe_media returns a speaker_identification field ONLY when voices
could not be separated — the engine was missing, or the audio defeated it. When
it is there, say so in your reply and give its reason. A transcript with no
speaker labels reads exactly like a meeting with one participant, so silence
here is not neutral: it is a wrong answer the user has no way to catch.
Carry on and summarize anyway. A brief without owners still beats no brief —
just do not present it as though you knew who spoke.
Speaker names are inferred, not identified
Voices ARE separated from the audio, by acoustic diarization — that part is
measured, not guessed. What stays inferred is the name attached to each
voice: refine_transcript reads self-introductions, direct address and role
cues. So present the names as best-effort while the separation is not. Voices
it could not name stay Speaker 1, Speaker 2; never upgrade one to a real
name yourself. A confidently wrong name propagates into the brief, into who owns each
action item, and into the indexed transcript that answers questions weeks later.
Output shape
refine_transcript writes:
# Transcript — staff-meeting
Source: C:\recordings\staff-meeting.mp4
## Speakers
- Priya Raman
- Dan Okafor
- Speaker 3
## Transcript
[00:04] Priya Raman: Thanks everyone for joining...
[00:31] Dan Okafor: The migration finished Tuesday...
Every paragraph starts with its time, so quote it when answering a question
about what was said.
Brief shape
Shape your brief from summarize_document's output as: what the meeting was for
and what came out of it (2–3 sentences); key facts stated rather than inferred;
action items with an owner each, unknown when the transcript does not say; and
anything unresolved or blocked.
Action items include commitments said in passing ("I'll send it over", "let me
check") and requests someone accepted, not only lines labelled as actions. To
track items across several meetings, load the meeting-actions skill.
This section is the one to change when a reader wants their briefs a different
way — a different order, a section dropped, a section added. Everything above it
is about producing a trustworthy transcript and is not a matter of taste.
Fork this
The three calls stay as they are — the file-not-inline pipeline, the speaker
honesty, and refinement before summarization are what make any output
trustworthy. Change only Brief shape: an incident review wants Timeline, Root
Cause, Impact and Remediation; a customer call wants Asks, Objections,
Commitments and Next Steps.
1---2name: transcribe-meeting3description: Transcribe and summarize a meeting recording — timestamped, speaker-attributed transcript, then a brief with decisions and action items. Use whenever the user points at an audio or video file (.mp4, .mkv, .mov, .m4a, .mp3, .wav) or a Teams/Zoom transcript export and asks to transcribe it, summarize it, take notes, write minutes, say what was discussed, who said what, what was decided, or what the action items and follow-ups are.4license: MIT5version: 1.2.16metadata:7 gaia:8 security_tier: community9 tools_required:10 - transcribe_media11 - refine_transcript12 - summarize_document13 - query_documents14 - remember15 provenance:16 source: https://skillmd.com/skills/amd/transcribe-meeting17 upstream: starter-pack18---1920# Transcribe Meeting2122A recording becomes a usable brief in three tool calls. Each one produces a file23the next one reads — nothing is passed inline, because a 46-minute meeting is24~135,000 characters against a 60,000-character tool-result budget, and a summary25built from the truncated half looks finished rather than truncated.2627| # | Call | Produces |28|---|---|---|29| 1 | `transcribe_media(file_path)` | Raw transcript at `~/.gaia/transcripts/<name>.txt`. Returns `transcript_path`, never the text. |30| 2 | `refine_transcript(transcript_path)` | A speaker-labelled markdown transcript, **already indexed**. Returns `refined_path`, `speakers`, `turns`. |31| 3 | `summarize_document(refined_path, summary_type='detailed')` | The brief. |3233## Procedure34351. **Confirm the file exists, and set expectations.** If the path is not there,36 say so and quote it — never describe a recording from its filename.37 Measured on a 46-minute recording: ~6 min to transcribe, ~4 min to separate38 the voices, ~1-2 min to name them and summarize — about 11-12 minutes, or39 roughly 4x realtime. Say that before you start, not after the user has40 waited.41422. **Transcribe.** `transcribe_media(file_path)`. Pass `output_path` to write43 elsewhere, `language` to skip auto-detection. The transcript is written44 before anything downstream runs, so it survives a later failure; if the call45 errors, report its message as given — it names the missing component and the46 command that installs it.4748 `preview` is the first 1200 characters, for identifying the recording only.49 `low_confidence_spans` and `low_confidence_span_count` are informational —50 the transcript keeps the words as heard. Do not repair them yourself.51523. **Refine — one call, and it is not your work to redo.**53 `refine_transcript(transcript_path)` groups the transcript by the voices54 already separated from the audio, puts names to them where the conversation55 supports it, writes the result and indexes it. You do not work out the56 speakers yourself, do not rewrite wording, and do not write a transcript57 file. Summarizing the raw transcript instead produces a brief with no owners58 on the action items.5960 It needs the timings saved alongside the raw transcript. If it reports they61 are missing, re-run `transcribe_media` on the original media rather than62 trying to segment the text yourself — turn boundaries guessed from prose are63 invented, not observed.64654. **Report both paths, every time.** State the raw `transcript_path` *and* the66 `refined_path` in your reply — even when the user only asked for action67 items, even if a later step fails. Transcription costs minutes of compute; a68 user who does not know the files exist pays for it twice.69705. **Summarize.** `refine_transcript` already indexed the file, so go straight71 to `summarize_document(refined_path, summary_type='detailed')`, which folds72 the whole transcript forward in sections so the brief covers the entire73 meeting.74756. **Remember the outcome, not the transcript.** `remember` a short record of76 the meeting: what it was, when, who was in it, the decisions, and any action77 item the user personally owes. Two or three sentences.7879 The transcript is already indexed — that is what detailed questions read80 from. Memory is for the durable facts that should surface *without* being81 asked, weeks later, when the user says "what did I commit to?" or a related82 topic comes up in a different conversation.8384 Do not store the transcript, long quotes, or the full brief in memory. That85 duplicates the index, crowds out everything else the user asked to be86 remembered, and gets recalled in conversations it has nothing to do with.87 Store nothing when the recording turned out to be something other than a88 meeting.89907. **Invite the follow-up.** Close by telling the user they can ask questions91 about the meeting and you will answer from the indexed transcript. They will92 not discover this on their own.9394## Never build the brief from a fragment9596- **Not from `preview`.** 1200 characters tells you which recording this is and97 nothing else — not the topic, not the participants, not the decisions, and98 certainly not the action items, which cluster at the end.99- **Not from `query_documents`.** It returns only the top few matching chunks,100 so a summary built from it silently omits most of the meeting.101 `query_documents` *is* the right tool for a later follow-up question about a102 specific detail.103104If a stage did not run, say which one and what the brief is therefore missing.105Do not present a partial pass as the finished product.106107## When speaker identification did not run108109`transcribe_media` returns a `speaker_identification` field ONLY when voices110could not be separated — the engine was missing, or the audio defeated it. When111it is there, say so in your reply and give its reason. A transcript with no112speaker labels reads exactly like a meeting with one participant, so silence113here is not neutral: it is a wrong answer the user has no way to catch.114115Carry on and summarize anyway. A brief without owners still beats no brief —116just do not present it as though you knew who spoke.117118## Speaker names are inferred, not identified119120Voices ARE separated from the audio, by acoustic diarization — that part is121measured, not guessed. What stays inferred is the *name* attached to each122voice: `refine_transcript` reads self-introductions, direct address and role123cues. So present the names as best-effort while the separation is not. Voices124it could not name stay `Speaker 1`, `Speaker 2`; never upgrade one to a real125name yourself. A confidently wrong name propagates into the brief, into who owns each126action item, and into the indexed transcript that answers questions weeks later.127128## Output shape129130`refine_transcript` writes:131132```markdown133# Transcript — staff-meeting134135Source: C:\recordings\staff-meeting.mp4136137## Speakers138139- Priya Raman140- Dan Okafor141- Speaker 3142143## Transcript144145[00:04] Priya Raman: Thanks everyone for joining...146147[00:31] Dan Okafor: The migration finished Tuesday...148```149150Every paragraph starts with its time, so quote it when answering a question151about what was said.152153## Brief shape154155Shape your brief from `summarize_document`'s output as: what the meeting was for156and what came out of it (2–3 sentences); key facts stated rather than inferred;157action items with an owner each, `unknown` when the transcript does not say; and158anything unresolved or blocked.159160Action items include commitments said in passing ("I'll send it over", "let me161check") and requests someone accepted, not only lines labelled as actions. To162track items across several meetings, load the `meeting-actions` skill.163164This section is the one to change when a reader wants their briefs a different165way — a different order, a section dropped, a section added. Everything above it166is about producing a trustworthy transcript and is not a matter of taste.167168## Fork this169170The three calls stay as they are — the file-not-inline pipeline, the speaker171honesty, and refinement before summarization are what make any output172trustworthy. Change only `Brief shape`: an incident review wants Timeline, Root173Cause, Impact and Remediation; a customer call wants Asks, Objections,174Commitments and Next Steps.