--- name: meeting-transcriber description: Turn a meeting, sales call, customer interview, or discovery recording into structured notes — a clean transcript plus TL;DR, decisions, and an action-item table. Use whenever Dhruv asks to transcribe a call or recording, take notes from a meeting, pull action items or decisions out of a conversation, summarize an interview, or clean up a raw transcript — including when he just drags in an audio or video file and asks what's in it. For transcribing a call happening live right now, use the live-meeting-transcriber skill instead; its finished transcript comes back here for the notes step. --- # Meeting Transcriber Turn a recording into notes someone can act on without re-listening. Transcription is the cheap part; the structure added afterward is the product. **Runs from anywhere.** Scripts are referenced by absolute path, and the `meeting-transcriber` subagent is installed at user level (`~/.claude/agents/meeting-transcriber.md`, symlinked to this repo), so no `cd` is needed — recordings usually live on the Desktop, not in the repo. **Repo location.** Commands below reach the repo through `$REPO` so they work wherever it was cloned. Set it first in any shell that runs them: ```bash REPO="${MARKETING_AGENTS_ROOT:-$HOME/github/marketing-agents}" ``` If the clone lives anywhere other than `~/github/marketing-agents`, export `MARKETING_AGENTS_ROOT` once in `~/.zshrc` and every skill follows. ## Step 1 — Resolve the input file Do this before anything else, and never skip to transcription on a guessed path. - **If a path was given** (dragged in, pasted, quoted): use it verbatim. Do not retype it. - **If no path was given**: glob the likely locations rather than asking Claude to recall a filename — ```bash ls -lh ~/Desktop/*.{m4a,mp3,wav,mp4,mov} ~/Downloads/*.{m4a,mp3,wav,mp4,mov} 2>/dev/null ``` Then show the candidates with size and duration and **ask which one**. Never pick between similarly-named recordings on your own — screen recordings from the same session differ by a couple of characters and picking wrong costs a long extraction pass. ### Never retype a filename macOS screen recordings contain a **narrow no-break space (`U+202F`)** before "AM"/"PM". It renders identically to a normal space and is a different byte sequence — retyping the name produces `No such file or directory` on a file that plainly exists. So: always pass paths through from a glob or from what the user supplied, always quote them, and use tab-completion or `ls` output rather than transcription of a displayed name. If a file "doesn't exist" but you can see it, this is why — recover with a glob: ```bash ls -b ~/Desktop/*.mov # -b reveals the escape sequences ``` ## Step 2 — Triage size before transcribing Groq documents a **25 MB** upload cap, but **treat ~15 MB as the real ceiling** — see below. Recordings routinely run into the gigabytes, so check first: ```bash ffprobe -v error -show_entries format=duration -of csv=p=0 "" ``` | Situation | Action | |---|---| | Audio file already under ~15 MB | Transcribe directly | | Video (`.mp4`/`.mov`/…), any size | Extract audio first — always | | Audio over ~15 MB | Re-encode to 16 kHz mono **at 32k** before transcribing | | `ffprobe` returns nothing | The file may be truncated or still being written. Say so; don't proceed | **Encode at 32k. Not 48k.** 32k ≈ 0.23 MB/min, so ~60 min lands near 13 MB and a 2-hour recording still fits: ```bash ffmpeg -i "" -vn -ac 1 -ar 16000 -c:a aac -b:a 32k ".m4a" ``` Whisper is speech-recognition, not audiophile playback — 32k mono at 16 kHz costs nothing in transcription accuracy and buys a large margin against the ceiling. ### Why ~15 MB, not 25 MB Measured on 2026-07-25 against a 57-minute recording: | Size | Result | |---|---| | 358 KB (60 s clip) | ✅ transcribed | | 13 MB (32k, full) | ✅ transcribed | | 21.5 MB (48k, full) | ❌ `[Errno 32] Broken pipe` mid-upload | | 42 MB (original) | ❌ over cap | 21.5 MB is comfortably under Groq's documented 25 MB and still failed — the connection dropped while sending the body, which points at a network/proxy body limit rather than Groq rejecting it. The API itself was reachable and the key valid throughout (verified separately). So the documented cap is not the binding constraint; **encode for ~15 MB and the question doesn't arise.** A broken pipe on upload is therefore a *size* symptom, not an auth or connectivity one. Re-encode smaller before touching the key or the network. Write extracted audio to the scratch area or alongside the source — never into `Claude Outputs/`, which is for deliverables. **Never chunk a file silently to get under the ceiling.** If a recording won't fit even at 32k (roughly 2 hours+), say so and offer the local fallback (Spokenly does real diarization and has no cap, but this agent does not drive it). ## Step 3 — Get a transcript - **A transcript already exists** (`.txt`/`.md`/`.vtt`/`.srt`, or pasted text) → skip transcription entirely, go to Step 4. - **The meeting was also captured in Granola** → use the Granola connector for context only: `list_meetings` or `query_granola_meetings` to find it, `get_meetings` for its notes and attendee names (load with ToolSearch, keyword `granola`). That gives Step 4 real names for speaker attribution. Granola does not return full transcripts on his plan, so the transcript still comes from the recording or a pasted export. - **Live / in-progress** ("transcribe as I go", a call happening now) → that's the **`live-meeting-transcriber`** skill (`agents/live-meeting-transcriber/`). It needs the Mac's microphone, so it only runs in Claude Code on the machine — hand off to it; its finalized `transcript.txt` comes back to this skill's Step 4 for speaker attribution and notes. - **A finished recording** → ```bash REPO="${MARKETING_AGENTS_ROOT:-$HOME/github/marketing-agents}" python3 "$REPO/agents/meeting-transcriber/scripts/transcribe.py" "