Hi all!
I wanted to share the bigger setup of my project management workflow using PM Codex, since a few pieces might be useful to people here even if you never touch PM Codex.
The problem
I have 1–3 hours of meetings a day. They’re bilingual, full of jargon, and people talk over each other. I record them on a Plaud Note Pro and Plaud Desktop, but I didn’t want Plaud’s subscription because I'm cheap. My local GPU was too weak to transcribe well, so I built my own pipeline in the cloud.
Plaud Note Pro
raw audio
│
▼
Claude scheduled task
finds new recordings, submits them
│
▼
Cloud Transcriber (Modal)
diarize · transcribe · name speakers
│
▼
Auricle (Claude)
clean up · confirm speakers · draft MoM
│
▼
PM Codex
change request to be reviewed
1. Plaud: raw audio only
The official Plaud connector fetches the recording file. I don’t use any of Plaud’s paid transcription or summary features.
2. Claude scheduled task: hands-free pickup
A Claude scheduled task runs in the cloud every hour during working hours on weekdays, so my computer doesn’t need to be on. Each run:
- Looks through my Plaud recordings from the last few days.
- Skips anything too short to be a meeting, anything unusually long, and anything that’s already been sent.
- Sends the remaining recordings to the transcriber and doesn’t wait for them to finish.
- Cleans up old transcription jobs.
- Reports back in one to three lines.
It’s deliberately simple. It never asks questions, never touches PM Codex, and never guesses who’s speaking. It runs on a small, fast Claude model with automatic approval, so it costs almost nothing. By the time I sit down to review a meeting, the transcript is usually already waiting.
3. Cloud Transcriber: diarized ASR on Modal
It runs on serverless GPUs that are billed per second and scale to zero. For each recording:
- pyannote works out who spoke when.
- Whisper large-v3, Qwen3-ASR and Qwen3-Omni (on vLLM) each transcribe every speaker turn.
- An ensemble keeps, for each turn, the version the engines agree on most. It flags likely hallucinations instead of silently deleting them.
- Voiceprints name the speakers. A speaker is named automatically only when the match is strong and clearly ahead of the next candidate. Otherwise it asks me.
It costs about $0.25 per audio hour, which fits inside Modal’s free monthly credit for my usage. Before each job it checks how much credit is left and switches to a free fallback if there isn’t enough.
4. Auricle: the workflow layer (a Claude skill)
- When I say “Run auricle for a meeting at 2PM”, Claude picks up the finished transcript, or starts one if the scheduled task hasn’t picked up that recording yet.
- It cleans up the transcript using a glossary and the attendee list pulled from PM Codex. This fixes misheard names and acronyms.
- It asks me only about speakers it couldn’t identify, and shows example quotes so I can tell who’s who.
- It learns from my answers: confirmed speakers become voiceprints, and corrected terms go into the glossary.
5. Into PM Codex
Auricle doesn’t write anything directly. It files one change request through the Claude bridge with the minutes, new tasks, risks and timeline updates. I review it and sign off in the app like any other change.
What I learned
- Context beats model size. Giving the engines and the cleanup pass a glossary and attendee list did more than swapping engines. In my tests, the cleanup pass alone cut word errors by 3–6 points and recovered 20–35% more key terms.
- Combining engines beats any single one. My ensemble beat every engine on its own, and it also beat Plaud’s built-in transcript on the same audio.
- Crowded rooms are still hard. Every engine struggles with who said what when people talk over each other. Voiceprints are good for naming speakers but don’t fix that.
- Voiceprints are biometric data. If you build something similar, check your local privacy rules and make it easy for people to delete theirs.
The transcriber and the Auricle skill aren’t public yet, but I’m happy to answer questions about the approach. If you’ve wired PM Codex into your own capture or transcription tools, I’d love to hear how!
Hi all!
I wanted to share the bigger setup of my project management workflow using PM Codex, since a few pieces might be useful to people here even if you never touch PM Codex.
The problem
I have 1–3 hours of meetings a day. They’re bilingual, full of jargon, and people talk over each other. I record them on a Plaud Note Pro and Plaud Desktop, but I didn’t want Plaud’s subscription because I'm cheap. My local GPU was too weak to transcribe well, so I built my own pipeline in the cloud.
1. Plaud: raw audio only
The official Plaud connector fetches the recording file. I don’t use any of Plaud’s paid transcription or summary features.
2. Claude scheduled task: hands-free pickup
A Claude scheduled task runs in the cloud every hour during working hours on weekdays, so my computer doesn’t need to be on. Each run:
It’s deliberately simple. It never asks questions, never touches PM Codex, and never guesses who’s speaking. It runs on a small, fast Claude model with automatic approval, so it costs almost nothing. By the time I sit down to review a meeting, the transcript is usually already waiting.
3. Cloud Transcriber: diarized ASR on Modal
It runs on serverless GPUs that are billed per second and scale to zero. For each recording:
It costs about $0.25 per audio hour, which fits inside Modal’s free monthly credit for my usage. Before each job it checks how much credit is left and switches to a free fallback if there isn’t enough.
4. Auricle: the workflow layer (a Claude skill)
5. Into PM Codex
Auricle doesn’t write anything directly. It files one change request through the Claude bridge with the minutes, new tasks, risks and timeline updates. I review it and sign off in the app like any other change.
What I learned
The transcriber and the Auricle skill aren’t public yet, but I’m happy to answer questions about the approach. If you’ve wired PM Codex into your own capture or transcription tools, I’d love to hear how!