The idea
I wanted a feature series with drama scenes, thirty minutes per episode, with me as narrator. The title was already set: “Nostalgie und KI” (Nostalgia and AI).
Each episode takes an object from the past and follows it to the present. The answering machine leads to the cloned voice, the Walkman to the AI companion. The assistant suggested the subtitle, “Jede Zukunft hat eine Vergangenheit” (Every future has a past), and I kept it.
After a few clarifications I wrote “okay - go for it. full shot. no risk no fun.” My first direction on the first version: the scenes drag, one answering machine is enough.
The format
The narrator is my clone, the ichbinsprecher.de brand voice, spoken through ElevenLabs Eleven v4. My first note on the voice was about its pressure. Out of eight test variants, one stage direction before every paragraph survived, [relaxed, warm, understated], which lowers the tone.
Most scenes are played by the Behrens family, and it is fictional. Gisela, her husband Werner, her son Thomas and his daughter Lina carry the series from 1972 to 2026. A “Fräulein vom Amt” (the switchboard operator of old German telephone exchanges) was also planned as a recurring framing character. In the second round of notes I wrote: “Fräulein raus… voll doof.” (Drop the operator… totally lame.) Since then the narrator gives the years.
Several episodes play the same situation twice, once then and once now.
After the title, every episode mentions in passing “Produziert mit KI.” (Produced with AI.), and the closing credits spell it out.
Eight episodes, 32 versions
| Episode | Versions | Length | What triggered the rounds |
|---|---|---|---|
| 1 · Answering machines and AI voices | 10 | 21:10 | 31 notes in the first review round, dedicated music, one sentence at the end |
| 2 · Pocket calculators and ChatGPT | 6 | 22:30 | v1 ran eleven minutes; second research pass, teachers, watermarks |
| 3 · Brockhaus (the German encyclopedia), Wikipedia and AI answers | 4 | 21:12 | context right after the scene, “ein Meter siebzig” (one metre seventy) |
| 4 · Video store, Netflix and AI films | 4 | 23:04 | two content requests: one more scene, TV on demand |
| 5 · Slide-show evenings and AI photos | 2 | 18:13 | three notes |
| 6 · Quelle, Neckermann (German mail-order houses) and AI agents | 2 | 17:48 | five notes, serious music for the Nazi-era history |
| 7 · Walkman, AI music and AI companions | 2 | 20:32 | 18 notes: Walkman, content warning, AI music |
| 8 · The millennium and AI forecasts | 2 | 18:43 | two notes |
Episode 1 needed ten versions, episodes 5 to 8 two each. In between lies a rulebook. After the first review I wrote: “Mir ist wichtig, dass ich mich dem perfekten Podcast nähere” (What matters to me is getting closer to the perfect podcast). The assistant turned my 31 notes on episode 1 into rules and a checklist that every new episode is checked against before my review.
My goal was thirty minutes; I signed off eighteen to twenty-three. Length comes from content, never from padding, says the rulebook.
The truth in the text
Language models tell stories fluently, even where they are wrong. So every script carries a source table, 340 entries in total.
Episodes 5 to 8 went through a fact-check before I heard a single line. It ran as checking tasks for the assistant, so-called agents. For each episode, two checked the facts against sources on the web and one checked the rulebook. Every factual finding was cross-checked by two more, which adds up to 196 agents. Among 560 checked factual claims they found 83 errors: 29 wrong, 41 overstated, 13 unsupported. The script for episode 6 said Neckermann’s first catalogue was from 1950. His textile mail order was already running in 1937/38, through Karl Joel’s “Aryanized” company.
Then the real recordings. Episode 1 plays recordings from 1860, 1901 and 1927. The idea of cleaning them up with AI was mine. The assistant built in a cleaned-up version of the 1860 recording. I listened and wrote: “Na - ganz offenbar hat die KI was neues erfunden hier… Wir streichen das Experiment” (Well, the AI has obviously invented something new here… We’re dropping the experiment). The rulebook has its own line for it. No experiments in which the AI invents something.
The review
For the review I asked for a dedicated page. A player on whose timeline I place notes. The assistant replies at the same spot. There were 93 notes, 92 implemented, one discarded.
In the big round I placed 62 notes across all eight episodes. 23 were about the music, mostly its fades. At some point I only wrote one word, “Blendenthematik” (the fade issue). 17 concerned pronunciation, eight breaths and edits, eight the pause before the closing credits, five the content. One was praise.
Eleven notes were about a single word. My clone said “Wok-Man”, a German reading of the word. At the third instance I wrote “OMG WALKMAN NICHT WOK-MAN” (OMG WALKMAN NOT WOK-MAN). Of the solutions the assistant offered, I chose saying it less often and checking strictly. The word now comes up nine times instead of sixteen. Every remaining instance has to sound English in the phoneme check, in at least seven out of eight measurement windows.
Breaths were another recurring theme. On an in-breath at the end of a take I noted: “Das wirkt wie ein Schnittfehler.” (That sounds like an editing error.) When I asked whether breaths can be detected, the assistant built a breath checker. It found all seven spots I had marked, and two I had missed. Across 682 take boundaries it raised no false alarm.
The last version of episode 1 was made for a single sentence. In the outlook, the assistant had cut a wrong year out of the take, and the sentence lost its ending. I heard it at the very end, nothing else. Six new takes were made for v10, chosen by measurements.
The check before I listen
One of my shortest briefs read “setz das alles um bitte vorher will ich nicht hören” (please implement all of it, I don’t want to listen before that). The production followed it. Before a version reaches the review page, a chain runs.
- Speech recognition compares every take with the script.
- A phoneme check listens for names and problem words such as Margenfeld and Walkman.
- An intensity check has outlier takes regenerated.
- The breath checker looks for in-breaths leading nowhere.
- A music checker measures the fades and holds −16 LUFS. All eight final versions pass it without errors.
The mixing software and checking programs in the shared library come to a little over 9,000 lines of code.
The music
All the music comes from Suno. Opener and closing credits come from my music bed set, plus seven series themes. Piano, vocals, brass and anything heroic are excluded.
I liked the sound; the fades felt clumsy. As a benchmark I named a documentary series on OpenAI by Deutschlandfunk (German public radio), and the assistant measured it internally. There the music sits in blocks of a little over a minute. In my versions it rose in every pause between sentences, 9 to 15 times a minute.
This led to new mixing software in 15 versions. I decided the direction blind, in six listening comparisons. Four times I chose differently from the proposal: one music bed per chapter instead of music with dry stretches, a more present bed, the first paragraph of “Was bleibt” (What remains) without music, and the old ending. The bed now sits 17 LU below the voice. The plan was 23; the reference suggested 20. In the final versions the music now rises only 0.7 to 2.3 times a minute.
Under the original recordings in the teaser, the music no longer fades out; it only dips.
Where the subject is death, grief, suicide or Nazi crimes, the music falls silent. For the Nazi-era history of the mail-order business I relaxed the rule: “wir dürfen schon Betten hier verwenden… halt kein lustiges.” (we can use beds here… just not a cheerful one.) Following my note, episode 7 carries a suicide notice in the opener, with the number of the Telefonseelsorge (the German crisis helpline).
Release and measurement
The recommendation was to launch with the trailer and three episodes. After my “GO GO GO” all eight went online at once. The AI notice is also in the feed in machine-readable form. On the episode pages, the scripts’ source entries are resolved into 570 individually linked references. Downloads are counted without cookies. The first count showed nine downloads, seven of them my own test requests.
What still sounds like amateur dramatics
About the scenes I once noted “etwas Laientheatermäßig” (a bit amateur-dramatics). The library voices can reach dubbing quality only to a limited degree, the assistant replied. I never reached the thirty minutes either. What holds the series together is something else. Its facts are cross-checked, its original recordings untouched.
A look at the case
Can a machine tell history in a way that makes you believe it? After this season I would say: if it distrusts itself. Agents found the factual errors, and a cross-check confirmed every single one. What no check knew about, the ear found. A Wok-Man, one breath too many, a fade too short. Each of these notes became a rule, some became a checking program.
At API price, concept, scripts and fact-check cost about €344; sound and polish afterwards about €1,334. The truth was cheaper than the sound. What I actually paid for the assistant was about €62, because it ran on the plan; with voices and images, about €89.
What this means for an organization
Strategy. The first draft has become cheap. About four fifths of the assistant cost up to sign-off came after it. A budget measured against the first draft would have been almost five times too small here.
Role. It takes a cross-check, and a machine can do it. The rules it checks against are set by a human. Here there were two: every factual finding is cross-checked independently, and nothing is invented, not even to improve something. The second came from an idea of my own that I dropped.
Process. Note, rule, checking program. The notes on episode 1 became a rulebook, the breath notes a checking program. What the ear has found once, the machine looks for next time. All 159 mixes in the production folders are still there, including the discarded ones.
What stays human. The limits. Operator out, AI restoration out, content warning in, serious music for the Nazi-era history, nothing goes out without asking. After the first review I wanted to get closer to the perfect podcast. Three versions later I wrote “folge 1 nahezu perfekt” (episode 1 nearly perfect). The “nearly” is something a human hears.