Content Lab

From an idea to a measurable experiment

Content decisions often rely on a few headline numbers. I wanted a more useful picture: what was said, how the episode opened, why people responded, whether they kept watching, and which production choices were worth testing again. Content Lab connects those questions with the actual scripts, recordings, transcripts and audience feedback.

The working loop is research → hypothesis → script → voiced prototype → recording and editing → distribution → analysis. I keep the evidence and editorial decisions together so the next episode starts with what I have learned.

Writing and listening before recording

I write episode scripts and use LLM-assisted research to develop angles, opening hooks, questions and contrasting perspectives. Candidate scripts can be assessed against the account’s previous topics and performance before committing to a recording.

ElevenLabs turns the dialogue into voiced prototypes. The pipeline parses speaker turns, generates individual audio segments and assembles them with FFmpeg, including pauses between lines. Listening to a draft helps me test whether the opening is clear, the dialogue sounds natural, the pacing works and the argument develops coherently. Optional vertical previews add visuals and subtitles.

These are internal production experiments. I revise the writing after listening, while audience-response hypotheses are assessed separately against published episodes and their recorded performance.

Account analytics at several levels

A history of the account, not just its latest totals

Ingestion adapters collect video metadata, descriptions, publication times, comments and transcripts. Timestamped snapshots preserve changes in views, likes, comments, shares and saves. That history supports per-video growth curves and comparisons across the account. Transcript collection is incremental, with local Whisper as a fallback when a transcript is unavailable.

Content, engagement and retention

I analyse episodes by topic, opening-hook type and emotional framing, alongside their transcripts and duration. The metrics include likes, comments, shares and saves per 1,000 views, early comment velocity and median performance by content category. Imported Studio data adds average watch time, completion, retention and followers gained where available.

These views answer different questions: an episode may reach many people, spark a discussion, attract followers or become something viewers save. Keeping those signals separate makes editorial choices more informative than a single engagement score.

Audience questions and the wider creator landscape

Comment analysis helps surface recurring questions, disagreements and possible follow-up topics. Reply drafts use the episode transcript and conversation context for review. A separate discovery workflow maintains a creator watchlist and leaderboard. Outlier scores compare each video with its own creator’s median performance, making unusual results visible across accounts of different sizes.

Testing hypotheses and keeping the counterexamples

I use the system to investigate questions about subject matter, opening lines, framing, length and visual presentation. Video labels, normalized metrics, transcript analysis and LLM-generated assessments provide inputs to that investigation. Portable HTML reports with SVG charts make the comparisons inspectable.

The workflow also keeps an insight ledger and an evolving editorial playbook. A promising pattern can become a candidate for another episode; a contradictory result can change the hypothesis. I account for factors such as topic differences, video age and production changes, and separate my own seeded comments from organic audience signals when interpreting a discussion.

These are observational comparisons and practical production tests. Their value is in generating better questions and decisions without mistaking a successful episode for proof that one specific change caused its performance.

Editing and republishing across formats

The production workflow takes an existing recording through transcription, edit proposals, rendering and review. Word-level ASR supplies candidate edit points, while audio-energy analysis helps find actual silence. Structured cut proposals are checked for valid boundaries, ordering and duration before FFmpeg renders them. I review whether the result preserves complete thoughts, the opening and the ending.

Repurposing adapts the material, titles, descriptions and vertical-video presentation for short-form destinations such as TikTok, Reels and YouTube Shorts. The implemented automated upload path prepares YouTube Shorts after review and uploads them privately. Platform-specific publication remains a deliberate decision.

The delivery system tracks previously handled material, persists job state and binds approval to the exact preview, file and metadata version. Hash checks detect changes after approval. Upload receipts and duplicate checks help prevent an interrupted job from publishing the same material twice.

Architecture: models propose, code checks, I decide

Python adapters feed normalized SQLite records and timestamped measurement snapshots. Analysis operates on stored evidence, while transcripts connect the analytics to what was actually said. This separation lets me change a research question or scoring method without collecting everything again.

I chose SQLite and portable HTML/SVG reports for a compact, inspectable stack. Language models support research, script development, interpretation and edit proposals. Deterministic code handles calculations, media constraints, file integrity and delivery state. Human review covers the parts that need listening, editorial judgment and publication approval.

Applied example: two women exploring money, technology and everyday life

I co-create and co-host a Polish-language conversational channel built around two women bringing their perspectives to questions that connect personal experience with wider social change. We discuss money and financial choices, AI and technology, social pressures and relationships, alongside topics closely connected with women’s lives, including beauty expectations, motherhood and the everyday decisions people rarely examine in depth.

The editorial approach starts with a recognisable human question and looks for the mechanisms behind it: economic incentives, changing norms, technology or the evidence behind a widely repeated claim. I develop research and episode scripts, check the supporting material and use the dialogue format to make the discussion accessible. Different perspectives create room for questioning assumptions and disagreement without reducing the audience to a stereotype.

Recorded audience results

The videos also generated 5,248 comments. A conversation about saving money versus enjoying life recorded 82,900 views and 203 comments; a discussion of financial expectations around relationships reached 112,200 views and 548 comments. These examples show how everyday subjects can open a broader conversation about money, values and social expectations.

A communication practice supported by engineering

I use Content Lab throughout this work: research and script development, ElevenLabs listening prototypes, editing, account analysis and repurposing. The analytical record contains 4,026 performance snapshots and 141 transcripts, so I can examine the material behind a result and revisit an editorial hypothesis as more evidence appears.

This is also a foundation for a future channel focused more directly on AI and engineering. It gives me practical experience explaining complex ideas to people outside my professional field, listening to their questions and judging whether an explanation is clear enough to be useful. The technical channel is a next step I am preparing for through both the production system and the communication practice.

Video metrics: latest stored snapshot per video, 16 October 2026. Views are cumulative video plays, not unique viewers. Follower figure: 5,313 in the stored account-profile snapshot of 14 October 2026. Each video is counted once in the totals.