Podcast Production Pipeline
Automate the workflow around an episode — guest research, outline, show notes, social promo — so you only spend time on recording.
Creative & Buildingadvanced~2h setup
- Tools
delegateweb_fetchmemoryexec- Channels
telegramdesktop- Uses
subagents
The conversation is 45 minutes. Everything around it — researching the guest, drafting an outline, writing show notes, posting to social — takes another 4–6 hours per episode. A pipeline collapses the prep and post-production work into ~30 minutes of human review.
What it does
- Guest dossier — when you book a guest, agent researches them (recent talks, papers, twitter, books) and produces a 1-page brief
- Outline draft — given the brief + the angle you want, drafts 10–15 questions ordered for natural arc
- Show notes generator — given the recording transcript, produces timestamped chapter markers, key quotes, and links mentioned
- Social promo set — 3 pull-quote cards (X/LinkedIn) and a 30-second reel script
What you'll need
- Web fetch for research
- Delegate for parallel subagents
- Memory for episode state
- Audio transcription (Whisper-cloud or your own — feeds into the pipeline as text)
Setup
1. Define the episode workflow
Send to Flowly
Standard episode workflow. Memory tag pattern: "episode:<slug>:*"
When I say "new episode with <name>":
1. Guest dossier — spawn researcher subagent:
"Research <name>. Return:
- Bio (2 paragraphs, focus on what's relevant for an audio audience)
- 3 talking points they tend to repeat (find via interviews/talks)
- 2 contrarian or under-discussed views they hold
- 5 candidate angles for a 45-minute conversation
- Cite all sources"
Save to "episode:<slug>:dossier", send to me.
2. After I confirm an angle, spawn outline subagent:
"Draft an outline of 10-15 questions following [angle]. Format:
opening hook (their story), middle development (the real
conversation), closing (what should listeners do/think)."
Save to "episode:<slug>:outline".
When I say "process episode <slug>" with a transcript:
3. Spawn show-notes subagent:
"Read the transcript. Output:
- 5-7 chapter markers with timestamps and 3-word descriptions
- Top 5 quotable lines with timestamps
- All books, papers, products, names mentioned (linked where possible)"
Save to "episode:<slug>:shownotes".
4. Spawn social subagent:
"Generate 3 pull-quote cards (one per platform: X, LinkedIn, IG).
Each is one quote + 1-line context. Plus a 30-second reel script
pulling the most provocative moment."
2. Run it
Send to Flowly
new episode with Andrej Karpathy
Within 5 minutes: dossier delivered. You pick an angle:
Send to Flowly
Angle: how he thinks about LLM agents now vs his earlier RL work
Outline arrives in 2 minutes. You record the conversation. After recording:
Send to Flowly
process episode karpathy-2026-05 — transcript at /tmp/transcript.txt
Show notes + social set in your inbox before the audio is even uploaded.
Tips
- Dossier hits and misses. Public figures = great research. Less public guests = thin output; supplement with whatever they sent you.
- Outline is a starting point, not a script. Your real job during the conversation is to listen and follow what's interesting, not to march through 15 questions.
- Show notes need fact-checking. AI summaries hallucinate references. Quickly verify each book/paper before publishing.
- Reuse across episodes. The dossier from episode 12 becomes context for episode 47 if you re-book the guest. Tag aggressively.
- Don't auto-publish. Pipeline produces drafts. You publish. Always.