Turn one radio show into a week of content with AI

A two-hour live show is the most expensive thing a small station makes and the least reused. The fix is not “try harder at social” — it’s recording the show and treating that one audio file as the source everything else is derived from. This guide gives you the actual pipeline, with the copy-pasteable prompts, and is honest about the parts AI must not touch.

1. The insight: one file, then everything downstream

Add up what a live show costs you. Prep the night before, an hour of setup, two hours on air, the guest you spent a fortnight chasing. Then it ends, the stream moves on to automation, and the entire thing evaporates. Somebody who was doing the school run at 4pm has no way of ever hearing it. That is the real problem — not reach, not the algorithm. The show simply stops existing the moment it stops playing.

RadioKing has a widely-shared post listing 20 ways to repurpose your radio content. The list is fine. The problem with a list of 20 ideas is that you do three of them once, in a burst of enthusiasm, and then never again, because there is no order to them and no obvious place to start. What you need is not more ideas. You need one repeatable sequence you can run every week without deciding anything.

Here is the whole insight, and it is unglamorous: recording is the unlock. Not AI. AI is what makes the derived work cheap enough to bother with, but it has nothing to work on unless the show exists as a file. A station that records every show and does nothing clever has more raw material than a station that reads every AI-tools listicle and never presses record.

2. The pipeline, end to end

Eight steps. Run them in this order, every week, for the same show. The order matters: each step consumes the output of the one before it, so skipping ahead means doing work twice.

1. 1Record the show. One file, the whole show, uncut. Nothing downstream exists without this.
2. 2Transcribe it. Machine transcription with timestamps. This is the step that converts audio into something a model — and a search engine — can actually read.
3. 3Mine the transcript for the 3–5 genuinely good moments. Not the informative moments. The ones where something happened.
4. 4Cut the clips at the in and out points the transcript gave you. This is the only step that stays slow, and the only one worth your hands.
5. 5Write show notes and chapters from the transcript. Draft with AI, then edit — always in that order, never the reverse.
6. 6Publish the episode to your podcast feed, with the notes attached, so the show has a permanent address.
7. 7Draft the newsletter from the same transcript, in your own voice, pointing at the episode.
8. 8Schedule the social posts — the clips plus quote captions — spread across the following week, not dumped in one afternoon.

Once it’s a habit, the whole run is 60–90 minutes for a two-hour show, and about 40 minutes of that is cutting video clips by hand. Everything else is drafting and one editing pass.

3. Recording: the step everything depends on

If you take one thing from this guide: turn recording on and leave it on. Record shows you think are unremarkable. You are a bad judge of your own show in the hour after it airs, and the moment that travels is almost never the one you planned.

On Nobex, show recording comes in from the Pro plan ($49/mo) and includes publishing straight to a podcast RSS feed — the full walkthrough is in record shows and publish a podcast. That matters more than it sounds, because the RSS feed is what turns a recording into something with a URL, an episode number and a place on Apple Podcasts and Spotify. A file sitting in a folder is not content. A file with a permanent address is. Podcast hosting on Nobex is usage-billed and off by default, so the feed costs nothing until people actually download episodes.

Two practical recording habits that make the AI steps dramatically better. First, have your guest say their own name and what they do, on mic, near the top — transcription models get names wrong constantly, and one clean pronunciation gives you something to correct against. Second, leave a beat of silence around the good bits. When you feel a moment land, don’t talk over the tail of it. Clean edges mean a clip you can cut in thirty seconds instead of five minutes.

4. Prompt craft: the part that decides whether any of this is good

This is where the difference lives. “Summarise this transcript” gets you the same beige paragraph everyone else gets, because you asked for the average of everything ever written. Useful prompts do three things: they define quality in observable terms, they demand verbatim material rather than paraphrase, and they give the model permission to return less. That last one matters more than you’d think — a model asked for five things will invent the fifth.

Finding the moments worth cutting

You are a radio producer reviewing a transcript of my show.

Find the 5 strongest 30-90 second moments. A strong moment is one where:
- someone says something they clearly had not planned to say
- a specific detail appears: a number, a name, a place, a date
- the emotional register changes: laughter, a long pause, an admission
- a question gets an answer that surprises the person who asked it

Do NOT choose moments that are merely informative, or well phrased,
or that summarise the episode.

For each moment, give me:
1. start and end timestamp
2. the first sentence of the clip, verbatim
3. the last sentence of the clip, verbatim
4. one sentence on why it works, addressed to me, not to a listener

Rank them strongest first. If fewer than 5 qualify, return fewer
and say so.

TRANSCRIPT:
[paste]

Why it’s shaped this way. A model left to itself picks the most *articulate* passage, which is almost always the least interesting one, so the criteria describe observable events rather than asking for “the best bits”. The verbatim first and last sentences are your in and out points — you search for that string in your editor rather than scrubbing a two-hour timeline. And the closing line is doing real work: without explicit permission to return fewer, you get five moments from a show that had two.

Show notes that don’t announce themselves as AI

Write show notes for this episode, using only the transcript below.

RULES
- Open with one sentence stating what actually happened. Not what the
  episode "explores", "dives into" or "takes a look at".
- Then 4-6 bullets. Every bullet must contain something specific from
  the transcript: a name, a number, a place, a claim. Test each bullet:
  if it would still be true of a different episode of this show, it has
  failed. Delete it and write another.
- Then a chapter list: timestamp + a 3-6 word label. Label the content,
  not the format. "Why he sold the van", not "Interview continues".
- Banned words: dive, delve, journey, unpack, explore, insights,
  landscape, tapestry, testament, resonate, elevate, seamless.
- No sentence may begin with "In this episode".
- Add nothing that is not in the transcript. If you are unsure whether
  something was said, leave it out and note it at the end.

TRANSCRIPT:
[paste]

The banned-word list is not a style preference, it’s a filter. Those words are the statistical fingerprint of generated text, and readers now recognise them at a glance even if they can’t say why. The interchangeability test on bullets is the important rule though: generic bullets are the actual failure mode, and “would this be true of a different episode?” is a test the model can apply to its own output. The final instruction — flag rather than guess — turns hallucinations into a checklist instead of a landmine.

A newsletter that sounds like you

I am giving you three things: a VOICE SAMPLE, a TRANSCRIPT, and a TASK.

VOICE SAMPLE (500-800 words I wrote myself, past newsletters or scripts):
[paste]

TRANSCRIPT (this week's show):
[paste]

TASK: Write this week's newsletter, 250-350 words, as if I wrote it.

Match the voice sample on: average sentence length, how often I use
contractions, whether I address the reader as "you", how I open, how I
sign off, and whether I use humour. Copy the habits, not the content.

Do not use a word that does not appear somewhere in the voice sample,
unless the transcript requires it (names, places, song titles).

Structure: one paragraph on the moment from the show worth caring about,
one paragraph on what is coming next week, one line linking the episode.

Afterwards, list separately every sentence you were least confident
about and why.

Telling a model to write “in a warm, conversational tone” gets everyone the identical warm conversational tone, because that adjective maps to one place in the model. Voice transfer needs a sample, not a description. The vocabulary constraint is the single most effective anti-AI-tell available to you: it stops the model reaching for its own favourite words and forces it back into yours. And the confidence list at the end is your edit queue — start there, and you’ll usually find the two sentences that would have embarrassed you.

Captions that quote the guest instead of describing the episode

From the transcript below, write 5 social captions.

Each caption MUST be built around a verbatim quote from the transcript,
in quotation marks, attributed by name. The quote does the work. Your
words only frame it.

- Maximum 2 sentences of your own around the quote.
- Never describe the episode. "We talked about X" is banned.
- Never end on a question unless the guest asked one.
- No hashtags, no emoji, no "link in bio".
- Give the timestamp of each quote so I can verify it against the audio.
- If a quote needs trimming, use an ellipsis. Never change a word
  inside the quotation marks, including filler words and grammar
  mistakes. If you cannot quote it exactly, do not use it.

TRANSCRIPT:
[paste]

Descriptive captions ask a stranger to trust you before you’ve given them anything. A quote *is* the thing — it works on someone who has never heard of your station. The instruction not to clean up grammar inside quotation marks is the safety rail: a model’s instinct is to tidy speech into prose, and a tidied quote is a fabricated quote with your guest’s name on it. The timestamp requirement exists so you can check in ten seconds rather than trusting on faith. More on what actually gets a station in front of new people in how to get radio listeners.

5. What you must never publish unedited

The pipeline above is fast enough that you will be tempted to remove the human from it. Don’t. Three categories are non-negotiable, and they’re the three that carry real consequences:

• Anything attributed to a guest. Models paraphrase and smooth by default. A quote that is 95% right is 100% wrong: you have put words in a real person’s mouth and published them under your station’s name. Check every quote against the audio at the timestamp. Every one, every time.
• Any factual claim. Dates, figures, chart positions, release years, company names, job titles, the spelling of anybody’s name. Transcription mangles proper nouns and models confidently repair the mangling into something plausible and wrong. If a number appears in your show notes, you personally verified it.
• Anything naming a person. Not just guests — callers, artists, local businesses, the councillor who came up in conversation. A generated sentence about a named individual is the fastest route to a correction, a complaint, or worse. Read every sentence containing a name as if the person will read it, because eventually one of them will.

A useful mental model: AI is a fast, tireless, slightly overconfident junior producer. You would not let that person publish under the station’s name without reading it. Same standard here, for the same reason.

6. What one show turns into

Concretely, from one recorded two-hour show. The effort column assumes the pipeline is a habit and the prompts above are saved somewhere you can paste them from:

The bottom row is the one people undervalue. A station with a year of published, transcribed shows has hundreds of pages of specific, genuinely unique text about its subject and its town. That is search visibility no amount of posting produces, and it accumulates whether or not anyone was listening live. It’s also the reason to put the episode on your own public page with the embeddable player rather than only on a social platform — the archive is an asset you keep.

7. Working out whether any of it works

The honest failure mode of a content pipeline is running it for six months without ever checking whether it moved anything. Two numbers are enough to start, and both are already in front of you.

One: live listeners in the hour after a clip goes out. Your listener analytics show sessions over time, so post a clip and watch that hour against the same hour the previous week. It’s a rough comparison, not a controlled experiment, but a clip that reliably moves nothing is telling you something real about the clip.

Two: which shows people actually finish. Podcast downloads tell you which episodes got picked up, and the pattern across a few months is the most useful programming feedback a small station gets. If interviews finish and music blocks don’t, that’s not a social-media insight — that’s a scheduling decision.

Give it eight weeks before you judge it. The pipeline compounds: the fourth month benefits from three months of archive, and nothing in the first fortnight tells you anything. What you’re watching for is direction, not a spike.

8. Questions people actually ask