A growing number of podcasters are skipping the dedicated podcast tools and building their own with Claude, ChatGPT, or a code editor. They write a few files, point an AI model at a transcript, and get show notes, clip candidates, and newsletter drafts in minutes.
The results are good enough that the obvious question is whether you should do the same. This post looks at what six people actually built, what a weekend of setup gets you, and the specific points where the do-it-yourself approach gets harder. It ends with a decision table so you can choose based on your own situation.
A caveat before we start. Nobody has published a benchmark of where these workflows break. What follows comes from what the builders wrote themselves, so it describes friction they mention and gaps their write-ups leave open, not measured failure rates.
- Podcasters build a folder per episode, a context file, a voice guide and one command per job.
- A weekend gets you notes, chapters and clips. Memory, voice, publishing and review time keep growing after that.
- The post ends with a build-or-buy table and a weekend plan.
What people have built
| Builder | What they made | What stayed manual or missing |
|---|---|---|
| Dr. Shantae Cofield, host of Prompting Curiosity | A Node.js command line tool built in Claude Code. npm run generate produces show notes and companion documents in about 2.5 minutes. npm run publish pushes formatted drafts to her blog, podcast page, and newsletter. Runs on a $20 a month Claude Pro plan. | The publish step creates drafts, so a final review remains hers. |
| Michael Simmons | A vault of about 11,000 notes, fed by Snipd and read by Claude Code to draft articles, clips, and summaries | The vault only works because somebody keeps filling it. |
| Didier Lopes, founder of OpenBB | An open-source Claude skill that turns a podcast video into a capped set of subtitled clips, with reasoning for each pick | Clip count is capped on purpose. |
| AI Business Systems | A VS Code and Claude Code template that turns one transcript into 8 outputs, guided by a voice-guide.md file | The voice guide is a file you fill in yourself. |
| Dan Vega, Spring Office Hours | Claude Code skills that edit raw audio, transcribe it, draft show notes, and open a pull request on his blog | He keeps the podcast host upload manual. |
| The Claude Cowork guide for podcasters | A walkthrough of one full episode, from transcript cleanup to clip candidates | No publish step, and it warns against automating too much too early. |
Sources: Cofield, Simmons, Lopes, AI Business Systems, Vega, Claude Cowork for Podcasters.
Cofield's summary of why she did it: "This is exactly what AI is for… building custom solutions to personal problems." Lopes explained his design choice by saying that nobody wants a link to a 90-minute episode with "trust me, minute 47 is awesome."
What a weekend gets you
Every one of these setups shares the same skeleton, and you can copy it in an afternoon.
- A folder per episode with the transcript in it.
- A context file the model reads first. In Claude Code it is
CLAUDE.md. It holds the show's name, format, host, and conventions. - A voice guide describing how the show sounds.
- One command per job: show notes, chapters, clips, newsletter.
Here is a starter CLAUDE.md you can adapt:
# Show context
Show: [name]. Format: [solo / interview]. Length: [minutes].
Audience: [who listens and what they want].
## Jobs
- show-notes: 120 word summary, 5 timestamped highlights, links mentioned.
- clips: propose 5 candidates, 30 to 60 seconds each, with the reason for each.
- social: 3 posts and 1 newsletter blurb.
## Rules
- Use only what is in the transcript. Never invent quotes, names, or numbers.
- Follow voice-guide.md for tone.
- Ask before publishing anything outside this folder.
## Known mistakes
- (Add a line each time the output gets something wrong.)
The numbers in the jobs section, a 120-word summary, five clip candidates, three posts and one newsletter blurb, come from the Cowork guide's worked example. Fixed limits are worth copying. A model told to "write a summary" writes whatever length it likes, and a model told to propose five clips gives you a manageable review.
Cost is low. Cofield's whole setup runs on a Claude Pro subscription with no separate API bill.
Where it stops being a weekend project
The friction points below come straight from these builders' write-ups. None is a dealbreaker. Each is a place where the work shifts from setup to upkeep.
Memory is a file you maintain
Every session starts fresh. The Cowork guide handles this with a CLAUDE.md that acts as a persistent instruction set, holding guest lists, formatting rules, and tone. That works, and it also means you are the one keeping it current. Simmons' vault shows the far end of this: 11,000 notes that Claude Code can read, all of which had to be captured first.
Voice is a separate job
The Cowork guide mentions a tone like "warm, plain spoken, curious" as an example inside its sample file, but it does not walk through building a voice guide. AI Business Systems' template includes one, and you write it. A model without a voice guide writes generic show notes, which is why it deserves its own setup. A follow-up post covers how to build one from your own episodes.
Publishing is where most builds stop
Look at the table again. Cofield built a publish command. Vega automated everything up to the upload and kept that step manual on purpose: "I still want to eyeball that upload before it goes live." The Cowork guide tells the model to ask before publishing anything outside the workspace. A draft in a folder is not a post on your site, so somebody still has to do the last step.
Review time does not disappear
Lopes caps his clip count deliberately, and his skill returns its reasoning for each pick so a person can judge it. Generating fifty clips is easy. Deciding which three to post is the work. Whatever you build, plan for a review step, because the model's output is a draft.
Upkeep is ongoing
Vega describes improving his skills by keeping a log of pitfalls and refining after each one. That is a sensible habit, and it is also a recurring task. Workflows that touch a website, a hosting platform, and a newsletter service break when any of those change.
Prompts alone are the lightest version
If a whole workflow feels like too much, prompt lists such as The Podcast Host's ChatGPT prompts and Castos's AI prompts are a fine place to start. They work for one-off jobs like drafting a title or an intro. They do not remember your show between sessions, so every prompt starts from zero, and you re-supply your context each time.
Build it or use a tool
| Build it yourself if | Use a tool if |
|---|---|
| You are comfortable in a terminal and enjoy tinkering | You want results without maintaining scripts |
| Your workflow is fixed and rarely changes | Your process changes with each season or format |
| You have one show and a small archive | You want the assistant to know every past episode |
| You want full control over every output | You would rather approve than assemble |
Episodemate is aimed at the right-hand column and is not released yet. If you fit the left-hand column, the builders above show it is a reasonable path.
A weekend plan
- Saturday morning: collect five transcripts and write your
CLAUDE.md. - Saturday afternoon: write a voice guide from those transcripts and run your first show-notes job.
- Sunday morning: compare the output with your last five hand-written show notes. Add every difference to the Known mistakes list.
- Sunday afternoon: add one more job, such as clip candidates, and set a fixed limit for it.
- Leave publishing manual until the drafts stop surprising you.
Checklist
- Transcripts stored one folder per episode
-
CLAUDE.mdwritten with fixed output limits - Voice guide written from your own words
- A Known mistakes list you update after every miss
- Publishing left manual until output is reliable
- A weekly review slot on your calendar