Podcast Production

Podcast production with generative tools

This skill covers producing a finished, publishable podcast episode when some or all of the audio is machine-generated. It is provider-neutral: it names TTS engines, music libraries, and hosts only as illustrative options, never as the method. The craft — format design, writing for the ear, voice direction, structure, rights, loudness, delivery, disclosure, and QA — is the same regardless of which tool renders the audio.

An episode is done when it (1) serves a defined show and audience, (2) sounds intentional and consistent, (3) uses only audio you have the right to distribute, (4) meets the loudness and file specs of its target platforms, (5) ships correct RSS metadata plus a transcript, (6) carries any legally or platform-required AI disclosure, and (7) has passed a listen-through QA. Skipping any of these is a defect, not a shortcut.

---

1. Scope and activation

Use this skill when the task is a spoken-word audio program: a topic explainer show, an interview series, a daily/weekly news brief, a narrative/documentary episode, a two-host conversational show, an internal briefing feed, or a show bible/format design for any of those.

Do not use it for:

• Music production — songs, beds, or stings as the deliverable (that is a music/audio-generation

concern). This skill consumes music and treats it as a rights and loudness problem.

• A single TTS line — an IVR prompt, a notification, one voiceover clip with no show context.
• Video where the picture leads — if the deliverable is a video and audio is a track under it,

the video workflow owns pacing and delivery. This skill applies when audio is the product.

If a request is ambiguous ("make an audio version of this article"), confirm whether the output is a one-off narration or an episode of a show. The two need different structure and metadata.

---

2. Format design (do this before writing a word)

The single most consequential early decision is format, because format dictates script style, voice count, structure, and length. Common podcast formats and what each demands (production heuristic, drawn from standard podcast-production practice — see Sources):

| Format | Typical length | Voices | Script style | Best for | |---|---|---|---|---| | Two-host conversational | 20–60 min | 2 (chemistry critical) | Outlined, not fully written | Ongoing relationship with audience; opinion/commentary | | Interview | 25–75 min | Host + guest(s) | Host questions scripted; answers live | Expertise, guests, evergreen back-catalog | | Solo / monologue | 5–30 min | 1 | Fully or mostly scripted | Teaching, essays, focused explainers | | Narrative / documentary | 20–45 min | Narrator + tape/characters | Fully scripted, structured in acts | Storytelling, high production value | | News brief / daily | 3–10 min | 1–2 | Tightly scripted, dense | Recurring, time-sensitive, habit-forming |

For a fully-synthetic show (every voice is TTS), the interview and two-host formats are the hardest to make convincing because they depend on turn-taking realism and chemistry; solo, narrative-narrator, and news-brief formats are the most forgiving. Weigh this when advising a format for a synthetic show.

Design the show once, in a short show bible, before designing episodes:

• Premise and audience — one sentence: who it is for and the promise each episode keeps.
• Cadence and length — a length band the format supports and a realistic release schedule.
• Voice identity — which voice(s), their names/personas, and a locked voice configuration so

episode N sounds like episode 1 (see §4).

• Signature elements — the intro line, the outro/call-to-action, the music theme, and any

recurring segments. These are the show's "sonic branding" and must be reused verbatim.

• Disclosure stance — decided up front (see §6), because it affects the script and metadata.

---

3. Writing for the ear

Scripts for audio are not text documents read aloud. The ear has no scrollbar, cannot re-read, and loses complex clauses. Rewrite prose into speech (production heuristics, widely taught in podcast scriptwriting — see Sources):

• Short sentences, one idea each. Break compound sentences. The listener holds only the last

clause in memory.

• Contractions and spoken register. "It's / you'll / here's," not "it is / you will / here

is." Write how a competent host actually talks.

• Front-load the point. Say the conclusion, then support it. Never bury the payoff behind a

long subordinate clause.

• Signpost transitions aloud. "Three things. First… / But here's the catch… / So what does

that mean?" These are the audio equivalent of headings.

• Kill the unpronounceable. Spell out or rephrase acronyms, symbols, URLs, and numbers.

"twenty-sixteen L-U-F-S," not "-16 LUFS." "dot com slash join," not "/join."

• Read it out loud. Anything that trips a human reader will trip a synthetic voice too and

will bore a listener. For a synthetic show, read-aloud testing is doubly important because you cannot ad-lib a save in post.

How much to script depends on format. Fully script the high-risk moments regardless of format: the hook/cold open, segment transitions, sponsor reads, and the close. For conversational and interview shows, outline the middle so the talk stays natural. For solo, narrative, and synthetic shows, script fully — a synthetic voice reads exactly what you give it, so the script is the performance.

When the script is the performance (fully-synthetic), also write performance direction into the script: mark intended pauses, emphasis, and emotional tone per paragraph, because those become your TTS instructions in §4.

---

4. Multi-voice casting and TTS direction

This is the craft that separates a convincing synthetic episode from an obviously robotic one. The goal is voices that are distinct from each other, consistent across episodes, correctly pronounced, and — for dialogue — plausibly reactive to each other.

4.1 Casting and consistency

• Lock voice configuration per persona. Save the exact voice identifier plus every generation

parameter (stability/consistency, similarity, style/expressiveness, speed) as the persona's "voice lock." Reuse it for every episode. Voice drift — the same nominal voice sounding different across episodes or even across a long file — is the top continuity failure in synthetic shows (practitioner observation, corroborated across TTS vendor guidance — see Sources).

• Cast for contrast. In a two-voice show, pick voices that differ in pitch, pace, and timbre

so listeners can tell who is speaking without name tags. Two similar voices are worse than one.

• Segment long scripts by speaker/role and generate each speaker's lines in that speaker's

locked config, then assemble. This preserves consistency better than switching voices mid-request on engines that support only one primary voice per call.

• Higher stability/consistency settings reduce drift and expressiveness together. For a

narrator that must sound identical for 30 minutes, bias toward stability. For a character that needs range, accept more variation and regenerate takes until consistent (heuristic).

4.2 Pronunciation control

• Fix names, jargon, and foreign words explicitly. Do not hope the model guesses. Use the

engine's pronunciation mechanism: phoneme markup (IPA or CMU Arpabet via SSML `), a custom pronunciation/lexicon dictionary, or, as a last resort, phonetic respelling in the text ("KAI-roh" for "Cairo"). Documented fact: SSML ` supports IPA and CMU Arpabet on major engines; some newer expressive models honor a lexicon or respelling but not full SSML — verify per engine (verified 2026-07-10 against ElevenLabs, Google Cloud TTS, and Azure Speech docs).

• Build a per-show pronunciation glossary in the show bible (recurring names, the show title,

sponsor names) and apply it every episode so "the host's name" never changes pronunciation.

• Numbers, dates, units, and symbols are read inconsistently. Normalize them in the script

("July tenth, twenty-twenty-six"), which also helps human clarity.

4.3 Emotional pacing and performance

• Punctuation is your primary pacing tool. Periods = full stop; commas = short breath;

ellipses = a hanging pause; question marks lift the final pitch. Sentence length sets rhythm — short sentences accelerate, long ones slow down. Around 140–160 words per minute reads as natural, engaged speech (production heuristic — see Sources).

• Apply emotion at the paragraph/section level, not word-by-word. Set a tone for a passage;

use per-word emphasis sparingly. Stacking many emotion directions or switching every sentence produces unnatural tonal lurches (practitioner heuristic corroborated across TTS best-practice docs — see Sources).

• Use the engine's expressive controls deliberately: SSML `` (rate/pitch/volume),

` for engineered pauses, `, or an engine's audio/emotion tags. Where the engine is prompt-driven rather than SSML-driven, put the direction in a delivery note the engine reads. Verify which markup your chosen engine actually honors before authoring — support is uneven (verified 2026-07-10).

• Generate multiple takes and choose. Synthetic delivery is stochastic; the first take is

rarely the best. Budget for regeneration.

4.4 Host chemistry in generated dialogue

Chemistry between synthetic hosts must be engineered, because the voices are generated independently and do not actually react. Techniques (heuristics):

• Write the reactions in. Interruptions ("—wait, say that again"), back-channels ("mm-hm,"

"right,"), and callbacks to earlier lines. If it is not in the script, it is not in the audio.

• Vary turn length. Real conversation is uneven — a long point, a two-word reply. Scripting

even, equal turns is a tell.

• Control the seams in assembly. Turn-taking realism lives in the gaps. Tighten or overlap

the joins between speaker clips so the reply does not sound like it was recorded in a different room a week later. Consistent room tone under both voices sells co-presence.

• Do not fake a live remote. If you add fake "over the phone" filtering to sell realness, that

edges toward deception; keep it clearly a produced show.

---

5. Episode structure and assembly

5.1 Structure

A conventional episode spine (production heuristic; adapt to format):

1. Cold open (optional, 10–30 s) — the single strongest moment or hook, before any branding.

Earns the listen. Strong for narrative and social-clip-driven shows.

1. Intro / theme (5–20 s) — show name, host, one-line promise, over the theme music. Keep it

short and identical every episode (sonic branding).

1. Episode tease — what this episode delivers, in one or two sentences.
2. Body / segments — the content, broken into clearly transitioned segments. Signpost each

segment change with a spoken transition and, optionally, a short music sting.

1. Ad slots — see §5.3. Mark them structurally so they can be inserted/removed or

dynamically served.

1. Outro (15–30 s) — recap, call to action (subscribe/share/link), sign-off, theme out.

Narrative shows use an act structure inside the body (setup → complication/turn → resolution or open question) rather than flat segments.

5.2 Assembly and editing

• Level the dialogue first, then place music and SFX under it. Speech intelligibility wins over

music every time.

• Duck music under speech (sidechain or manual automation) so beds sit roughly 12–18 dB below

the voice during talk and come up in the gaps (heuristic; tune by ear).

• Trim synthetic artifacts — clipped word-onsets, unnatural breaths, and swallowed

syllables are common in TTS. Cut or regenerate the offending line.

• Consistent room tone / silence between clips. Dead-digital silence between assembled TTS

clips sounds unnatural; a low consistent floor reads as one continuous recording.

• Match levels across segments so the intro, body, and ads are not wildly different volumes

before you do the final loudness pass (§7).

5.3 Ad slots

• Baked-in vs. dynamic. Baked-in ads are part of the file forever; dynamically inserted ads are

stitched at request time by the host and can be updated or removed. If the show will run ads long term, structure the timeline with clean, silent insertion points so ads can be dynamic.

• Disclosure carries into ads. A synthetic-voice ad read, or an AI-generated endorsement, may

trigger the same disclosure and consent duties as the show (see §6), plus advertising-law rules about endorsements. Never synthesize a real person appearing to endorse something without rights.

---

6. Fully-synthetic vs. hybrid, and the disclosure duty

6.1 The production decision

• Fully-synthetic — every voice is TTS. Cheapest and fastest to iterate; best for solo,

narrative-narrator, and news-brief formats; weakest for spontaneity and true interview dynamics.

• Hybrid — recorded human voice(s) plus synthetic segments (e.g., a real host with a synthetic

co-host, synthetic narration around recorded interview tape, or synthetic reconstruction of unavailable audio). Best of both, but multiplies the rights and disclosure surface: every recorded human needs a release, and every synthetic voice needs a licensed/consented source.

Choose hybrid when the show needs a real human's authority, spontaneity, or an actual guest, and fully-synthetic when scale, consistency, or cost dominate and the format tolerates it.

6.2 Disclosure and consent obligations

Treat these as requirements, not style choices. They are the highest-risk part of a synthetic podcast.

• Voice cloning requires documented consent from the voice owner. Cloning or replicating a

real person's voice without permission exposes you to right-of-publicity and voice-rights liability. As of 2026 multiple U.S. states protect voice as a distinct likeness right — Tennessee's ELVIS Act (effective 2024) is the first to name AI voice replicas explicitly, and California, Illinois, New York and others have related statutes; the federal TAKE IT DOWN Act (signed May 2025) targets nonconsensual synthetic depictions (documented fact, verified 2026-07-10 — see Sources; not legal advice — confirm current law for your jurisdiction and use). Consent must be specific, written, and documented; a verbal "yes" does not meet the bar. Never clone a public figure, a guest, or a co-host's voice without a signed release scoped to the use.

• Platform AI-disclosure policies apply to synthetic voices (documented facts, verified

2026-07-10):

• YouTube requires creators to disclose realistic altered or synthetic content at upload,

explicitly including synthetic/cloned voices and AI voiceovers that could mislead a viewer into thinking a real person spoke. A "How this content was made" label may be shown. Mass-produced, low-effort AI content risks demonetization under 2025 monetization updates.

• Spotify does not down-rank AI-assisted content per se but bans unauthorized voice clones,

aggressively removes spam/low-quality mass-produced audio, and is adopting a DDEX-based AI disclosure standard surfaced in-app. Its "Verified" program excludes profiles that primarily represent AI personas.

• Apple Podcasts does not currently mandate a generic "this is AI" label for synthetic voices

in the base RSS spec, but its content policies still prohibit impersonation and require rights to all audio. (Verify current policy at publish time — platform policy is volatile.)

• When in doubt, disclose. A brief, honest note ("voices in this episode are AI-generated" in

the show notes and/o

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: calesthio
• Source: calesthio/generative-media-skills
• License: MIT
• Homepage: https://github.com/calesthio/OpenMontage

Install and usage instructions live in the source repository linked above.