I don't edit this podcast. Here's who does.
A step-by-step guide to launching a podcast where you sit down, hit record, and everything after that gets handled. The exact machine, the prompts, and the skill file. Free.
TL;DR
Everything below, in the order it shows up.
The stack
Seven tools. One of them costs money. No editor, no paid host, no VA to get started.
Descript
Record, edit by deleting words from the transcript, captions, clips, and it hosts the share links. Its AI editor is called Underlord.
Creator planClaude
The operator. Connected to Descript, it runs the edit instructions, reads the transcript, writes everything, and drives the browser to publish.
SubscriptionSpotify for Creators
Free podcast host. Generates your RSS feed and distributes to Spotify.
FreeApple Podcasts Connect
One-time RSS submission. After that every episode flows from Spotify on its own.
FreeYouTube
Home for the long-form video, chapters, and the thumbnail.
FreeGoogle Drive
One folder per episode so nothing gets lost.
FreeIG · TikTok · Shorts · X · LinkedIn
Where the five clips and the posts land, two or three a week.
FreeOne-time setup
About two hours. Do it once. Your progress here saves in this browser.
Every episode: the 10-step loop
Click a step. Each one is tagged with who does it and whether it spends Descript credits.
Camera always on, even if you think it's an audio show. Every asset needs to be able to feed video. Name the project consistently, something like The Jump Ep. 7 - [short topic], so Claude can find it by name.
Solo? Hit record. Guest? Create a Descript recording room and send the link; it records each person on a separate local track, so their wifi can't wreck your audio.
"Run the podcast skill on [project name]." If it's a guest episode, say so and give the guest's name, because the posts come out in the guest's voice instead of yours.
Claude reads the raw transcript first and lists every retake. Then one instruction to Underlord: remove filler words, remove silences over about 1.5 seconds, delete the listed retakes keeping the last take of each, and add captions to the long-form only. I use Descript's "Modern Yellow" caption template, no waveform.
Giving Underlord the retake list up front works far better than asking it to find them.
The step everyone skips and the one that saves the most time. Underlord's cleanup reliably over-cuts: it deletes "I, I" stutters entirely (both words), drops the setup line of a story, or leaves two takes back to back. Claude re-exports the transcript, lists every problem, and fixes them in one batched corrective call.
Catching a mistake in the transcript costs nothing. Catching it after a 15-minute render costs the render.
Claude scans the transcript and names the five strongest standalone moments, surprising, emotional or contrarian, then tells Underlord exactly where to cut. Each clip is 30 to 90 seconds, opens on a hook line, and stands alone. Clips get no captions; add those per platform when you post.
Naming the moments produces dramatically better clips than letting the AI pick blind.
Fill the whole frame, no black bars, crop centered on the speaker's face. If you sit off-center on camera, tell it the focal point (mine lands around x = 0.4). Leave the long-form horizontal.
Render one clip first, grab a frame, and eyeball the crop before rendering all five.
From the transcript: 3 to 5 title options, show-notes description, YouTube title + description + timestamped chapters (from an SRT export after the edit so the times match the final cut) + 10 to 15 tags, 5 X posts, 5 LinkedIn posts, 1 short blog post, a caption for each clip, one pull-quote. All in your voice, all built from real quotes.
Guest episodes: 5 LinkedIn + 5 X in the guest's voice, assembled into a handoff doc with the recording link, transcript and clip descriptions.
Long-form as 1080p video, the main composition as audio, and each of the five clips. Every one returns a permanent share.descript.com link with a Download button. Those go in the Drive doc, not the giant expiring signed URLs.
Renders run one at a time per project: roughly 15 minutes for a 20-minute episode, 2 to 3 minutes per clip.
This is the gate. Watch the long-form cut before anything goes public. If you hand-trim anything in Descript after the automated edit, Claude re-checks the project, re-renders, and recomputes chapters, because a trim invalidates all of it.
Once you say go, Claude drives the browser through Spotify for Creators: New episode → upload the audio → title, description, season and episode number, episode type Full, explicit No → Review → publish date Now. You give the explicit "publish it." It clicks. Spotify lists it within a few hours and Apple picks it up from the RSS.
Claude creates the episode's Drive folder and fills 05 Copy & Transcript with the transcript, the copy/posts doc, a publishing record, and a media-links doc. You download the video, audio and clips from the share links into 01 through 03.
Then the long-form goes to YouTube with thumbnail, description and chapters, and the clips go out to IG, TikTok, Shorts and X two or three a week alongside the X and LinkedIn posts.
What one recording becomes
Gotchas I learned the hard way
Each of these cost me at least one episode's worth of frustration.
Letters go missing, punctuation survives. Bizarre.
Flip the "HTML" toggle (top right of the box) with a real mouse click and type the description as <p>…</p>. Verify the box shows raw source before typing.
The full-quality audio master won't go through.
Re-encode a mono web mp3 for the upload (about 80 kbps for a 12 to 16 minute episode, about 48 kbps for 25+), then swap the master in later from episode settings. It's voice; mono is fine.
It can bounce to Upload and drop the draft.
Drafts auto-save under Episodes. Reopen from the list. From Review, the bottom-left Back is safe.
"I, I think" becomes "think." Story setups vanish. Duplicate takes survive.
Give it the retake list from the raw transcript on the first pass and tell it to keep the last take of each. It can restore over-cut sentences ("restore the sentences where I say X") but not single words inside a bigger removed block; patch those by hand.
Firing a fix while a render runs wastes the render.
Cancel the running job first. Different projects can run in parallel.
Even with the timecode option on.
Export SRT to get real times for YouTube chapters, and do it after the final edit.
Not a bug, a feature.
Links never go stale, so the Drive doc stays right after a re-render.
Title, thumbnail and the first 30 seconds are what matter.
I've shipped episodes with a missing word at 4:14. Nobody has ever mentioned it. Perfectionism is friction wearing a nice outfit.
Batch and launch
Record 5 to 10 episodes before you publish anything. Run the skill on each. Upload them all to YouTube as Scheduled and to Spotify as scheduled or drafts, one per week. Publish episode 1 everywhere on day one with your best clip and a launch post, and let the schedule carry the rest. Keep a 2 to 3 episode buffer from then on.
The whole point of the system is that consistency stops depending on discipline. Lower the bar until stepping over it feels silly.
The guest offer
Because the package is repeatable, it's also the pitch.
Come talk with me for 30 minutes and you walk away with two weeks of content you don't have to make: the recording, your transcript, 5 short clips, and 5 LinkedIn and 5 X posts, all pulled from what you actually said. You edit nothing. I'll have it to you within a week.
You're not asking a favor. You're handing them a gift basket. The same pipeline is a sellable service if you want it to be: coaches, agents, anyone who knows they should be posting and hates doing it.
Copy-paste prompts
Three prompts that run the whole thing without the skill file. Swap the bracketed values.
Prompt 1 · your own episode
Run my podcast repurposing pipeline on my Descript project "[PROJECT NAME]". 1. Find the project (list_projects) and export its transcript. Read it and list every retake / restarted sentence you can see. 2. In ONE prompt_project_agent call, tell Underlord to: remove filler words, remove silences over 1.5s, delete the retakes you listed (keep the LAST take of each), and add captions to the long-form using the "[CAPTION TEMPLATE]" template with no waveform. 3. Export the transcript again and read it. List anything over-cut (missing sentence subjects, dropped story setups, duplicate takes, dead air). Fix all of it in ONE corrective prompt_project_agent call. 4. From the transcript, pick the 5 strongest standalone moments (surprising, emotional, contrarian). In ONE prompt_project_agent call, create a 30-90s clip for each, named, starting on a hook line, with NO captions, and reframe each to vertical 9:16 filling the frame, focal point on my face. Leave the main episode horizontal. 5. From the transcript write, in my voice: 3-5 title options, a show-notes description, a YouTube description with timestamped chapters (from an SRT export of the FINAL cut) and 10-15 tags, 5 X posts, 5 LinkedIn posts, 1 short blog post, a caption for each clip, and one pull-quote. 6. Publish the long-form as 1080p video, the audio, and each clip (publish_project) and give me the permanent share links. 7. Create the Drive folder "Ep NN - [Title]" with subfolders 01-05 and put the transcript, copy doc, publishing record and media-links doc in 05. 8. Stop. I'll watch the cut. Don't touch Spotify until I say "publish it." Only prompt_project_agent spends credits. Batch it. Do everything else freely.
Prompt 2 · guest episode / content package
Build a guest content package from my Descript project "[PROJECT NAME]" for my guest [GUEST NAME]. 1. Find the project and export the transcript. 2. One prompt_project_agent call: remove filler words, remove long silences, remove retakes (keep last take), add captions to the long-form only. 3. Re-export the transcript, read it, fix any over-cuts in one corrective call. 4. One prompt_project_agent call: create 5 uncaptioned vertical 9:16 clips from [GUEST NAME]'s strongest, most quotable moments (feature THEM, not me). 5. From the transcript, written in [GUEST NAME]'s own voice and phrasing: 5 LinkedIn posts based on things they specifically said, and 5 X posts. 6. Publish the long-form, audio and clips and collect the share links. 7. Assemble one handoff doc titled "[GUEST NAME] — Content Package": the recording link, transcript link, the 5 clip descriptions and links, the 5 LinkedIn posts and the 5 X posts. Save it as a file I can send them. Only use prompt_project_agent where AI editing is genuinely needed.
Prompt 3 · publish to Spotify (after you've watched the cut)
I watched the cut and I'm happy. First re-check the Descript project (updated time + duration) in case I trimmed anything; if I did, re-render the media, re-encode the audio and recompute the chapters before anything else. Then re-encode the audio export to a mono web mp3 under 10 MB. Open creators.spotify.com in the browser (I'll log in myself), go to my show "[SHOW NAME]" → Episodes → New episode. Upload the mp3. On Details: title (no "Ep N" in the title), description typed as HTML with the HTML toggle clicked ON, Explicit No, Promotional No, type Full, Season [N], Episode [N]. On Review set publish date to Now, then STOP and show me. Publish only when I say "publish it."
The skill file
Save this in Claude as podcast-repurpose. Then every episode is one sentence.
Swap [CAPTION TEMPLATE], [SHOW NAME], [HOST], the focal point, and the thumbnail line. Everything else is exactly how mine runs.
---
name: podcast-repurpose
description: >-
Turn a Descript podcast recording into a full content package: edited video,
vertical clips, written posts (X, LinkedIn, blog), launch metadata, a Drive
archive, and the episode drafted on Spotify for Creators. Two modes: my OWN
episodes (posts in my voice) and GUEST episodes (posts in the guest's voice
plus a handoff doc). Trigger on "repurpose this episode," "run the podcast
skill," "make the content package," "clip this," "guest package for [name],"
"publish this episode," or any time a Descript project needs to become clips,
posts and a published episode.
---
# Podcast Repurpose
Turn one Descript recording into a full content package with minimal lift and
minimal AI-credit spend.
- **Self mode**: my own episodes. Written content in my voice.
- **Guest mode**: an interview. Written content in the GUEST's voice, assembled
into a handoff doc. If unclear, ask: "Your own episode, or a guest interview?"
## The one cost rule
Of the Descript tools, ONLY `prompt_project_agent` (Underlord) spends AI
credits. `list_projects`, `get_project`, `import_media`, `export_transcript`,
`publish_project` and the job tools are free. Batch Underlord instructions into
as few calls as possible (one cleanup, one corrective, one clips+reframe).
Never loop Underlord for things the transcript already gives you.
## Workflow
### 1. Locate or ingest
`list_projects` to find it by name, confirm with `get_project`. External file
or URL: `import_media` (auto-transcribes); confirm with `wait_for_job`.
### 2. Cleanup edit: ONE Underlord call
Export the raw transcript first and list every retake / false start. Then one
`prompt_project_agent` call: remove filler words, remove silences over ~1.5s,
delete the listed retakes keeping the LAST take of each, and add captions to
the LONG-FORM ONLY using the "[CAPTION TEMPLATE, e.g. Modern Yellow]" template
with no waveform. Never caption the clips.
### 2.5 ALWAYS re-read the transcript after every cleanup pass
`export_transcript` is free. Read it. Underlord regularly deletes "I, I" /
"we, we" stutters entirely (leaving no subject), drops story setups, keeps a
false start over the good take, or leaves duplicate takes. Fix everything in
ONE corrective call. Underlord CAN restore over-cut sentences ("restore the
sentences where [HOST] says X") but NOT single words inside a larger removed
block; flag those for a manual patch. Catching it here costs nothing; catching
it after a render costs the render.
### 3. Clips: ONE Underlord call
From the transcript, pick the 5 strongest standalone moments (surprising,
emotional, contrarian, highly quotable). Name each and point Underlord at the
exact lines. Each clip 30-90s, opens on a hook line, stands alone. Self mode:
my best moments. Guest mode: the GUEST's best moments. State explicitly:
"do not add captions to these clips."
### 3.5 Aspect ratio (same call)
Reframe all 5 clips to vertical 9:16 (1080x1920): fill the frame, no black
bars, crop centered on the speaker's face (focal x ≈ [0.4] if the host sits
off-center). Leave the main episode horizontal 16:9. Render ONE clip first,
pull a frame with ffmpeg, check the crop, then render the rest. Note: the
connector can't toggle Descript's "Center Active Speaker" smooth-follow; the
static smart crop is the default. Offer the manual toggle as an upgrade.
### 4. Export the transcript (free)
This is the raw material for all written content. For chapter timestamps,
export `format: "srt"` AFTER the final edit (the text formats don't render
timecodes).
### 5. Write the posts from what was ACTUALLY said
Real quotes and specifics, never generic filler.
- Self mode (my voice, use my voice skill): 5 X posts, 5 LinkedIn posts (one
idea each), 1 short blog post.
- Guest mode (the guest's voice, mirror their phrasing): 5 LinkedIn, 5 X.
### 5.5 Launch metadata (self mode)
3-5 title options (curiosity-forward + light SEO), show-notes description,
YouTube description (2-3 sentence hook, timestamped chapters, CTA, links),
10-15 tags, a caption/hook per clip, one pull-quote. Thumbnail: [your method,
or "skip, a designer makes it"].
### 6. Export the media (free)
`publish_project` Video 1080p on the main composition (long-form), Audio on
the main composition (for Spotify), and Video on each clip. Each returns a
permanent share.descript.com link with a Download button. Use those links, not
the expiring signed URLs. One publish/agent job per project at a time;
`cancel_job` before firing a fix. ~15 min for a 20-min 1080p render, ~2.5 min
per clip.
### 6.5 Publish to Spotify for Creators
GATE: the host watches the finished long-form and says "publish it" first. If
they say they watched/finalized it, re-check `get_project` (updated time +
duration) because a hand-trim invalidates renders, the web mp3 and chapters.
Prep: re-encode the audio to a mono web mp3 under 10 MB (~80 kbps mono for
<16 min, ~48 kbps for 25+). The host can swap in the hi-fi master later.
Drive the browser: creators.spotify.com (host logs in; never type a password)
→ show "[SHOW NAME]" → Episodes → New episode. Upload the mp3 to the file
input. Details: title (no "Ep N"), description as HTML with the HTML toggle
clicked ON with a real mouse click (the rich editor garbles typed text), art
defaults to cover, Explicit No, Promotional No, type Full, Season + Episode
number. Review: publish date Now, then STOP and show the host. Click Publish
only on an explicit go. Apple picks it up from the RSS automatically.
Gotchas: don't use the wizard "Back" on Details (bounces to Upload); drafts
auto-save under Episodes; native selects need type-ahead.
### 6.6 Archive to Drive (every episode)
Parent folder "[SHOW NAME] - Podcast". Create `Ep NN - Title/` with
`01 Long-Form Video`, `02 Clips (Vertical)`, `03 Audio`, `04 Art`,
`05 Copy & Transcript`. Populate 05 with: full transcript, the Copy / Posts /
Captions doc, a Publishing Record, and a Media Links doc listing the permanent
share links. Binaries are too big to push via tools; the host downloads them
from the share links into 01-03. Deliver small files (web mp3, art) in chat.
### 7. Package the output
Self mode: long-form link, Spotify episode (drafted or live), 5 clip links +
captions, 5 X, 5 LinkedIn, 1 blog, title options, YouTube description +
chapters + tags. Guest mode: one doc titled "[GUEST NAME] — Content Package"
with the recording link, transcript link, 5 clip descriptions, 5 LinkedIn and
5 X posts, saved as a file. Then remind the host of the guest outreach line.
## Quality notes
- Ship at 80%. Title, thumbnail, first 30 seconds are what matter.
- The best clips are the surprising, emotional or contrarian lines.
- Camera-on, video-first, always.
- Cadence: weekly, keep a 2-3 episode buffer.