Episode 1 Companion Playbook

A short, published podcast pilot: your idea, your script, your (consented) cloned voice and/or a disclosed AI co-host, heard by real listeners, with a written plan for how you'll measure what they do.

A small or midsize business owner with real expertise and no studio. You already explain your business all day, one conversation at a time. A podcast records the explanation once and lets it keep working - and it opens a channel for customers to send you their real problems.

• 10-30 minutes of clean recordings of YOUR voice that you own and consent to use. (More is better: the tool we used recommends 30 minutes; we made it work with less - expect extra tuning if you're under.)
• One pilot topic and a few quiet hours spread over a week.
• A computer and a small tool budget (see section 9).
• Willingness to listen to yourself a lot.

• Voice cloning (the path this show actually used): ElevenLabs (https://try.elevenlabs.io/rkr9n8d9uckb - affiliate link: the show earns a commission if you sign up through it, at no extra cost to you; the tool was chosen before any affiliate link existed), Creator plan - $12 for the first month, about $24/month with tax after that. Key settings used on the show: model "Eleven v3", Stability "Natural", 1.05x tempo on export. CRITICAL: turn OFF the "train on my data / model improvement" setting BEFORE uploading any recordings.
• Co-host voice (optional): pick from the tool's stock voices. Disclose the AI co-host in the episode itself.
• Free alternatives exist (open-source engines such as Piper). Honest note from our testing: we tried the free routes first and they did not pass the "would a human believe this is a human" bar for us. Your ears may differ - test before you pay.
• UnresolvedRecording your source audio: any quiet room and a decent mic or phone. ⟦QUESTION: no specific mic recommendation is grounded yet.⟧
• UnresolvedAssembly/editing: ⟦QUESTION: the show's editor is not yet chosen - any basic audio editor that can join clips works; placeholder pending the show's own pick.⟧
• UnresolvedHosting/distribution (Spotify, Apple, YouTube): ⟦QUESTION: distribution is not set up for this show yet; no steps here are verified. Placeholder - do not treat as instructions.⟧
• SHOW-NOTES LINKS (show rule, Andrew 2026-09-21): every episode's show notes link the actual tools used. Where a legitimate affiliate program exists, the show notes use Andrew's own affiliate link with clear disclosure; otherwise the plain official link. Tools are never chosen because of an affiliate program. This guide is linked from the episode's show notes and from buildingbackwards.ai.

1. 1Write the purpose in one sentence: what the show is for and who it's for.
2. 2Pick ONE useful first topic - the explanation you find yourself repeating most.
3. 3Write a short pilot script (aim 600-1,200 spoken words). Read it out loud. Rewrite it as text until it earns existing. Nothing gets voiced that hasn't been read.
4. 4Decide your evidence BEFORE publishing: 2-4 observable behaviors (see section 7). Write them down.
5. 5In your voice tool: turn off training-on-your-data first, then upload your recordings in small chunks.
6. 6Generate a SHORT test clip - one paragraph, not an episode.
7. 7Listen. Adjust pace/tone/wording. Re-render. Repeat 3-5 loops until it passes the bar: would a human believe this is a human? (Expect the first versions to be flat. That's normal. Keep looping.)
8. 8Pick the co-host voice (if using one) by the same bar, and write its disclosure line.
9. 9Render the full script in small chunks; stitch them; fix seams; cut what drags.
10. 10Fact-check every claim in the episode against something real.
11. 11Add the AI disclosure - in the episode and in the show notes.
12. 12A human (you) listens to every second. Only after that yes does it go out. Everything before it is drafts.

Put the pilot in front of 3-5 honest listeners who match your real audience - customers or peers, not family being supportive. Give them the actual audio, not a description. Watch what they do; don't interview them about what they'd do.

Pick 2-4 observable signals, written down in advance:

• Completion: do they finish the pilot?
• First-minute retention: do they bail in the first 60 seconds?
• Unprompted forwards or subscribes.
• Replies with real problems or topic requests (this show uses a dedicated email for exactly this).

People saying nice things is appreciated - but not proof of value. The ladder that counts: do they listen, do they come back, do they share it.

Decide keep / change / kill per signal. Change ONE variable per loop (topic, length, cold open, pace) so you know what moved the number. Rebuild and run it again. The first version isn't the answer - it's the first honest measurement.

• Tool cost: $12 first month, ~$24/month after (tax included) on the path this show used.
• Time: this show's first prototype went from idea to testable audio in days, with most loops taking an afternoon. Your mileage depends on your source recordings and your bar. Treat "days, not quarters" as the shape, not a promise.

• Consent: only clone a voice you own or have explicit permission to use. No exceptions.
• Privacy: training opt-out BEFORE any upload. Assume anything you upload leaves your machine.
• Disclosure: AI voices get disclosed, in-show and in notes, every episode.
• Facts: no claim airs that you can't back. Invented testimonials or metrics kill the trust this format builds.
• Where being wrong is genuinely expensive - money, safety, regulated claims, promises to customers - do the careful human work up front. That is real risk, not a fake test.

• Voice rights: cloning anyone else's voice, or using a voice across platforms - get legal advice first.
• Distribution at scale, paid promotion, or anything touching customer data/payments: bring in someone who has done it before.
• If the pilot becomes real infrastructure (sponsors, customer commitments), treat it like a product, not a side project.