Running a podcast is two activities: the conversation and the production. The conversation is irreplaceable human work. The production — guest research, prep questions, editing, transcription, clip extraction, social posts, show notes — is where AI compresses days of work to hours.
Pre-interview research
For each guest, AI compresses background research from 2-3 hours to 20 minutes:
Research [guest name] for a podcast interview about [topic].
Give me:
- 5 underexplored aspects of their work that interviewers usually miss
- 3 contrarian positions they hold
- 5 specific projects/papers/talks worth referencing
- Any recent (last 6 months) news involving them
- 3 angles they haven't been asked about in 10+ recent interviews
The last bullet is gold. Generic research surfaces what's been asked already; AI is decent at finding angles other interviewers missed.
Verify everything before using. AI hallucinates publication dates, paper titles, and sometimes conflates two people with similar names.
Question generation
Draft 20 interview questions for [guest] about [topic].
Mix: opening warmups, technical depth, contrarian/challenging, personal/origin, future-looking.
Avoid generic questions like "how did you get started."
For each question, suggest one follow-up that would push deeper.
AI will generate ~20 questions. You'll keep maybe 7-10 for actual use. The volume helps you find the angles you wouldn't have thought of.
Don't read questions verbatim during interview. Use them as a memory aid; let the conversation breathe.
Recording and transcription
For recording: standard tools (Zencastr, Riverside, SquadCast). AI doesn't help here.
For transcription: Whisper-based tools (Descript, Riverside, AssemblyAI) produce ~95% accurate transcripts in minutes. Cleanup is fast for English; harder for Chinese-language podcasts but improving.
Riverside's automated transcript with speaker labeling is currently the simplest. For more editing power, Descript treats the transcript as the editing interface — delete words to delete audio.
Clip extraction
This is where AI provides genuine leverage:
From this transcript, identify the 5 best 60-90 second clips for social media.
For each, give me:
- The timestamp range
- Why this clip works (clear hook, surprising claim, emotional moment)
- A suggested social post caption
- Suggested platform (X, LinkedIn, TikTok, YouTube Shorts)
The model surfaces hooks you'd miss reading linearly. Verify each clip works as a standalone — sometimes context that's clear in conversation isn't clear when extracted.
Tools that automate this: Opus Clip, Riverside Magic Clips, Descript's Composition. They're decent but the AI-curated clips often miss the human moments that go viral. Treat the AI's suggestions as one input among others.
Show notes and chapter generation
AI's structuring is genuinely useful here:
Generate show notes for this transcript. Include:
- 1-paragraph episode summary (warm, not corporate)
- Chapter timestamps with topic labels
- 5 key quotes worth highlighting
- 3 takeaways for listeners
- Resources mentioned (with URLs if findable)
- Guest links
Edit for voice. AI tends toward generic show-notes voice; your podcast probably has something more specific.
Social distribution
For each episode, you typically need:
- 1 X / Twitter announcement
- 3-5 X reply tweets quoting clips
- 1 LinkedIn post (longer, professional framing)
- 3-5 TikTok / Shorts clips with captions
- 1 newsletter blurb
- 1 IG carousel summarizing key takeaways
- 1 reddit post (if relevant subreddit)
AI generates first drafts of all of these in 20 minutes. Edit for voice; the structure is reusable.
When AI hurts
During the interview itself. Don't try AI-assist live. Reading from a screen kills conversation flow. The whole point of podcasting is the human exchange.
Trying to script the conversation. AI-suggested questions are fine; AI-suggested guest answers are useless because the guest will say what they say.
Auto-generating the entire post. AI clip-suggestions plus AI-written caption plus AI-cropped video produces generic content that audiences scroll past. The human craft of selecting and framing is what makes podcasts cut through.
Generating sponsor-read copy. Sponsor reads where the host's authentic voice matters; AI flattens this exactly when authenticity matters most.
A workflow timeline
Pre-interview (60-90 min, week before):
- AI guest research: 20 min
- AI question draft: 15 min
- Your own selection and prep: 30 min
During (60-90 min):
- Just the conversation. No AI.
Post-interview (90-120 min, day after):
- Transcription: automated, 5 min
- Editing: 60 min (mostly removing fillers, fixing audio)
- AI clip extraction + selection: 20 min
- AI show notes draft + edit: 20 min
Distribution (45 min):
- AI-drafted social variants per platform
- Your edits and personal angle additions
- Schedule everything
Total AI time saved per episode vs no-AI: roughly 5-8 hours.
When NOT to use AI heavily
For a podcast where the host's curation and editorial voice IS the differentiation. Conversations With Tyler-style podcasts succeed because Tyler Cowen specifically chose those questions. AI helps with logistics but should never substitute the editorial brain.
For very short podcasts (10-20 min) where the production overhead is small. The compression isn't worth the AI integration setup.
For podcasts in languages where AI tooling is weaker (Cantonese, Hokkien, regional variants). Quality drops; manual workflow may still be better.
Decision tree
- Pre-interview research and questions: AI heavy use, with verification
- During interview: no AI
- Editing for content quality: human first, AI tools for cleanup
- Clip extraction: AI suggestions + your selection
- Show notes / chapters: AI draft + your edit
- Social distribution: AI variants + your voice
Next steps
- Build a per-show prompt template you reuse
- Try AI clip extraction on your last 3 episodes; compare to what you originally clipped
- Set up automated transcription if you haven't
- Test whether your audience prefers AI-extracted clips vs your own selections — feedback can be surprising