The Guide
How Del Be Del is made
Every episode is one long conversation, recorded in a small room with three cameras and a microphone for each of us. What happens after the record button is a pipeline I built with code: the machine does the tedious, measurable parts and stops at checkpoints; I review, decide and make the edit by hand. This page shows the whole flow so other podcasters can borrow from it.
The goal of Del Be Del goes beyond this one show: I want it to be easy for anyone to start an interview podcast, without getting lost in the technicalities of websites and video editing, and to spend that energy on the human connection instead. Everything on this page is an effort to simplify that.
The one rule
Being wrong must create a bad suggestion, never a bad timeline.
Machine prepares → I decide → machine assists → I decide → machine verifies.
The flow
the machine
a human, by hand
back to the edit
Who answers what
Every question goes to the cheapest thing that can answer it
A model is used only where judgment is the job.
Tier 0
Code
Deterministic scripts: sync maths, loudness, colour measurement, frame counts. Same input, same answer, every time.
Tier 1
Local models
Speech-to-text and a small vision model on the Mac: "is the colour chart in this frame?", "who is speaking?"
Tier 2
The assistant
Runs the stages in order, reads the printed verdicts, writes the checkpoint reports. It never re-derives a number by eye.
Tier 3
The story editor
A stronger model, used once per episode, only for editorial judgment: what to cut, what to protect, the chapters, the shorts.
01Human
Record
One conversation, one sitting. Three cameras in a flat log profile: a close-up on the guest, a close-up on me, and a wide. One microphone per person into a mixer that records each voice on its own track, so the voices can be treated separately later. Before the first word, a colour chart is held at each seat so the cameras can be matched by measurement, not by eye, and a clap at the start of every segment gives the machine something to lock onto.
↑ Back to the flowWhat I do
- Frame three cameras, set them once, roll all three
- Chart pass at both seats, then held for the wide
- A clap per segment
- Talk, and forget the equipment
What the recording must carry
- Isolated voices, one track each
- A chart in every camera
- Log footage at the agreed settings; the ingest refuses anything else
02The machineCheckpoints 1–3
Prepare
Everything measurable happens here, unattended, and ends in a short report. The scripts identify which file is which camera, check the footage against the recording contract, lock every camera to the mixer's clock, clean each voice, and match the cameras to the chart. Each step prints what it measured and stops. I open the editor, check one thing per checkpoint, and say "continue".
↑ Back to the flowWhat runs
- Ingest: which file is which, contract check (blocking or warning), thumbnails to glance at
- Sync: audio correlation places every camera on the mixer's timeline; clap frames prove it; a stacked timeline is built; playback proxies render in the background
- Audio: per-voice stems, bleed removed using the transcript's speaker map, levels and loudness set, a compare timeline for A/B listening
- Colour: chart holds found from the transcript and a local vision model, white balance and exposure measured per camera, skin tones matched, one technical correction per camera plus one show look
What I decide
- Checkpoint 1: ten seconds of playback per segment, the clap sheet at a glance
- Checkpoint 2: raw versus prepared audio, one solo button
- Checkpoint 3: the reference chart, any proposed exposure trim, the look
03The machineCheckpoint 4
Suggest
The machine reads the whole conversation and proposes; it never cuts. A word-level transcript with speakers feeds a story pass that marks candidate cuts, protect zones for the silences that matter, and chapters. The isolated microphones answer "who is speaking" to the millisecond, so a camera proposal can be built from sound alone. All of it lands in the editor as markers and disabled, colour-coded clips on their own lanes: one click to judge each suggestion.
↑ Back to the flowWhat runs
- Transcript with speakers; silence and gap candidates
- The story editor's pass: cuts, protect zones, chapters, title options, the spoken intro script
- Camera proposal from the microphones; a review timeline with every suggestion as a marker or a lane
- In parallel, the shorts: the same story pass plans clips, code snaps every beat to spoken words, proves one voice per clip, cuts from the raw 4K, frames faces, adds captions, and re-transcribes the render to check it
What never happens
- No script cuts the episode
- A suggestion is never applied because the model is confident
- Titles and intros are options for me to pick from, never published as written
04Human
Edit
This is the part that stays human. I duplicate the proposal, play it, and decide suggestion by suggestion: cut it, keep it, or change the camera with one key. From this moment the timeline I touched is the source of truth. Every later step reads it; none rebuilds it from an old file, and none edits it in place. If a machine step needs to change something, it works on a copy.
↑ Back to the flowWhat I do
- Judge every suggestion in place
- Choose the camera where I disagree
- Protect the silences
- Name the chapter breaks, at most five
The guarantee
- A fingerprint of my timeline is recorded; if it changes, ownership is mine automatically
- No step ever modifies it; copies only
- Raw media is never written to
05The machineCheckpoints 5–7
Verify
The machine comes back to check my work, and to add the parts around it. Audio and picture QA list every doubtful join with a timecode and never auto-correct. The spoken intro, recorded to camera from a script, is synced, graded and levelled like the episode. Teaser, logo motion, music and the evergreen outro are assembled around the conversation. Every cut the machine places inside speech is proven by rendering it and transcribing the render. The export is marked unverified until loudness and a final check pass.
↑ Back to the flowWhat runs
- Draft render; audio QA and picture QA with timecodes and the exact fix for each
- Intro and outro: takes aligned to the script, graded from that shoot's chart, levelled to the episode
- Assembly: teaser, logo, intro, the edit, logo, outro; music cuts with fades baked in
- Chapter breaks placed on measured pauses, verified by a transcribed render
- Final check: loudness per part, silences, true peak, black frames; a delivery pass to the upload standard
What I decide
- Checkpoint 5: work the fix list, or ignore it
- Checkpoints 6–7: listen to the renders, trim the intro handles, pick the chapter breaks that ship
- The export starts only when I say so; the file is renamed "ready to upload" only after the checks pass
06Human and machine
Publish
Nothing on the way out is written from scratch. The YouTube text is assembled from the title I chose, the story pass, and chapter times derived from the final assembly. The transcript and captions come from the exported audio. Covers are built from the designer's templates. This website's episode page is generated from the podcast feed and that same transcript. The shorts go out on a schedule, and only after the full episode is live.
↑ Back to the flowWhat runs
- Title, description, chapters, links; transcript and captions from the final export
- Episode covers in every format from the layered templates
- Website page: episode, guest, transcript with timestamps, short links
- Shorts queue: one test post, then every other day across platforms
What I do
- Upload the file that says "ready"
- Approve the first scheduled short on each platform
- Keep the master, the mixer audio and the deliverables; let the raw footage go
Rules that never bend
What makes it safe to automate
Automation is only worth having if a mistake costs a minute, not an episode. These rules are the reason the machine can be trusted with the boring parts.
- Raw media is immutable.
- After I edit, the current timeline is the truth. Nothing is rebuilt from old analysis.
- A timeline I own is never modified; the machine works on a copy.
- Measured is not suggested; suggested is not applied; applied is not approved.
- Every step that changes something has a dry run.
- Uncertain means stop, report, and mark for review.
- A model's confidence is never an authorisation.
Chapters
The exact settings, room, cameras and audio chain, one decision per chapter: what's used, what was rejected, and why.
The first chapters are being written.