An AI-hosted podcast episode in under 8 hours
Samantha McLean
Samantha McLean turned the raw transcripts of two talks into a video podcast hosted by two fictional AI co-hosts, plus a short intro in her own voice. She reports the whole project took just under 8 hours from start to publish.
Story written by: the person who did it
The problem
Samantha McLean came back from a week of speaking events with a bad cold and could not record her planned podcast episodes on camera. She still wanted the episode made. So she set herself a challenge: build the whole episode with AI, using talks she had already given as the raw material.
How it was built
- 1Start from existing transcriptsPerson
- 2Ask Claude for the best anglesClaude
- 3Rewrite the script as two AI hostsClaude
- 4Generate the two voicesElevenLabs
- 5Split the audio and make avatar videosHeyGen
- 6Edit the episodeDescript
- 7Record her own introHeyGen
- 8Write the notes and publishClaude
Tool step Person step
1. Start from existing transcripts
She began with the raw transcripts of two talks she had already given, not from scratch. Her rule: AI is a multiplier, so you need something to multiply.
Check: Have ready: recordings or transcripts you own or may use.
2. Ask Claude for the best angles
She fed the transcripts into Claude and asked it to pick three takeaways her audience would care about most. This took about 10 minutes and gave her the structure for the episode.
ToolClaude
Check: Pick the angle with the most substance for your audience.
3. Rewrite the script as two AI hosts
A solo script in Claude felt predictable, so she asked Claude to rewrite it as a conversation between two fictional hosts, Sarah Chen and Marcus Rodriguez, who reveal partway through that they are AI.
ToolClaude
Check: Read the script and confirm it sounds like a real conversation, with the AI reveal made clear.
4. Generate the two voices
She created the two voices in ElevenLabs and split the script into four segments of about 5,000 characters each, which was the limit in her setup at the time. Marcus's voice kept drifting out of character unless Sarah's voice spoke first in a segment, so she ordered every segment that way; this step took about an hour including troubleshooting.
ToolElevenLabs
Check: Listen to every segment and confirm each voice stays in character.
5. Split the audio and make avatar videos
She made one audio track per host, with silence where the other host speaks, so each avatar in HeyGen only moves its mouth during its own lines. She generated six HeyGen videos, three per host, in about 30 minutes.
ToolHeyGen
Check: Watch each video and confirm the avatar only moves when that host is speaking.
6. Edit the episode
She assembled everything in Descript, cutting between the two hosts and adding b-roll made with Runway, Veo and clips from playphrase.me. This was the longest step at about 4 hours, because editing still needs human decisions.
ToolDescript
Check: Confirm the cuts and b-roll match what the hosts are saying.
7. Record her own intro
She asked Claude to script an intro explaining that the hosts are not real, going back and forth about 5 times to get the tone right, then recorded it on her iPhone. HeyGen turned a photo of her and that recording into the intro video.
ToolHeyGen
Check: Confirm the intro says clearly that the hosts are AI generated.
8. Write the notes and publish
She asked Claude for the YouTube description, email copy, show notes and social captions, which took about 20 minutes, then checked them against the finished episode. She uploaded the video to YouTube and Flightcast distributed it to Spotify, Apple Podcasts and other platforms from that one upload.
ToolClaude
Check: Check every description and caption against the actual episode before using it.
Before and after
Before
A week of speaking events left her with a bad cold and no way to record the podcast episodes she owed, because she did not want to be on camera.
After
A full video podcast episode with two AI co-hosts, published to YouTube, Spotify and Apple Podcasts, in just under 8 hours of work (self-reported). source
What made it work
- She started from transcripts that already existed, so the AI had real material to work from.
- She let Sarah's voice speak first in every segment, which kept Marcus's voice in character.
- She split the audio into one track per host, so each avatar only moved while that host was speaking.
- She checked Claude's descriptions and captions against the finished episode before using them.
- She budgeted human time where it was needed: editing took about 4 hours of her own decisions.
Follow along
Practice steps suggested by Ai Siftly, based on the account above. We have not run them ourselves.
Prepare the material
Have ready: one page of your own notes and one audience question. Remove private details before any AI step.
Example to copy: Audience: first-time buyers. Question: what should they prepare before a viewing? Approved facts: [checked notes].
You should get: A short approved source page.
Check: You own or may use the material.
Write the script
Ask an AI chat tool for a single-host opening, main points and close. Read it aloud, time it and shorten it to two minutes.
Example to copy: Write a two-minute single-host script using only these notes. Mark any claim not in the notes as [CHECK]. If using synthetic narration, disclose it. [notes]
You should get: A checked two-minute script.
Check: Resolve every CHECK; no invented numbers.
Record the words
Read the approved script into a phone or computer recorder. Or paste it into a text-to-speech tool whose voice licence and current cost you have checked. Listen to the whole file and fix mispronounced names, missing words or unwanted sounds.
You should get: A reviewed audio file and final script.
Check: Meaning matches the script, with no clipping or missing words.
Trim only if needed, then save
If you already know an audio editor, trim long pauses without removing meaning. If not, record a cleaner take instead of learning an editor for this exercise. Save the final file and play it in a separate player.
You should get: A two-minute local episode.
Check: Listen end to end; volume and length are right.
Optional: plan a second speaker
This is a separate extension, not needed to finish the exercise. Label every line HOST A or HOST B, and make a table showing who speaks, when each line starts and ends, and its audio filename. When combining tracks, leave silence on one track while the other speaker talks. Avatar animation and lip-sync need separate setup.
Example to copy: turn | speaker | start | end | audio filename
You should get: A speaker-labelled script and timing sheet.
Check: Do not assume past character limits, or that production time equals audio length.
Prepare notes for review
Draft the description and captions from the final script, then compare them against the exported audio before you think about publishing.
Example to copy: Write a factual two-sentence description from this final script; add no claims. [script]
You should get: Audio, transcript and checked description.
Check: Rights, disclosure and platform checks are complete; nothing publishes automatically.
Try it yourself
Find a matching workflow in the workflow finder. Search words: t u r n t a l k t r a n s c r i p t s i n t o A I p o d c a s t.
Sources and notes
- How I created a podcast episode entirely with AI (AI Powered Real Estate (Substack), 2025-10-25)
- All times, including the total of just under 8 hours, are self-reported by Samantha McLean. No audience, download or revenue results are given.
- Her setup reflects the tool versions she used in October 2025; an archive note on the page (September 2026) says the same. Current versions, limits and costs may differ.
- The roughly 5,000 characters per generation was the limit in her own ElevenLabs setup at the time, not a general limit. Check current ElevenLabs guidance.
- Use your own voice or a licensed synthetic voice. Never clone a voice without consent.
- The follow-along steps are Ai Siftly's suggested practice for a simpler two-minute audio exercise, not the subject's exact process.
- Source: Samantha McLean, 'How I created a podcast episode entirely with AI', AI Powered Real Estate (Substack), 25 October 2025, https://samanthamclean.substack.com/p/how-i-created-a-podcast-episode-entirely
Independent first-person. Page has an archive note (September 2026) saying tool steps reflect versions she used in October 2025. Prompts described, not quoted. Historical tool behaviour and commands are as documented at the source date; current versions may differ. Record dated 2026-10-27.